├── LICENSE ├── README.md ├── data_preprocessing.py ├── gat_lstm_model.py ├── testing.py └── training.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | **GAT-LSTM Model for Power Load Forecasting** 2 | 3 | This repository contains the implementation of the GAT-LSTM model, a hybrid approach that combines Graph Attention Networks (GAT) and Long Short-Term Memory Networks (LSTM) for short-term forecasting of power load. The model leverages the spatio-temporal dependencies in energy systems, incorporating graph-structured data (e.g., power grid topology) and temporal sequences (e.g., historical energy consumption and weather data). **Detailed explanation of the model and results are captured in the paper**: https://doi.org/10.48550/arXiv.2502.08376 4 | 5 | **Features** 6 | 7 | i. Edge Attribute Conditioning: Transforms edge features to influence GAT attention mechanisms effectively. 8 | 9 | ii. Graph Attention Network (GAT): Utilizes parallel GAT layers to capture spatial relationships and node-level interactions in a graph representing the power grid. 10 | 11 | iii. Spatial and Temporal Fusion: Combines the graph-derived embeddings with the sequence (temporal) data before feeding to the sequence processor. 12 | 13 | iv. LSTM: Processes the fused, temporally-aware embeddings to make the final hourly power load prediction. 14 | 15 | 16 | **Data Sources** 17 | 18 | i. Electricity (Load, PV, wind, etc.): https://gitlab.com/dlr-ve/esy/open-brazil-energy-data/open-brazilian-energy-data 19 | 20 | ii. Grid (Line length, capacity, efficiency, etc.): https://gitlab.com/dlr-ve/esy/open-brazil-energy-data/open-brazilian-energy-data 21 | 22 | iii. Weather (Air temperature, pressure, rainfall, etc.): https://www.kaggle.com/datasets/gregoryoliveira/brazil-weather-information-by-inmet?resource=download 23 | 24 | iv. Socio-economic (State-wise GDP): https://www.ibge.gov.br/en/statistics/economic/national-accounts/16855-regional-accounts-of-brazil.html 25 | 26 | v. Population (State-wise population): https://www.ibge.gov.br/en/statistics/social/population/18448-estimates-of-resident-population-for-municipalities-and-federation-units.html?edicao=28688 27 | 28 | 29 | 30 | -------------------------------------------------------------------------------- /data_preprocessing.py: -------------------------------------------------------------------------------- 1 | import os 2 | import logging 3 | import pandas as pd 4 | import numpy as np 5 | import torch 6 | import networkx as nx 7 | from sklearn.preprocessing import RobustScaler 8 | from torch_geometric.data import Data 9 | import matplotlib.pyplot as plt 10 | import pickle 11 | 12 | # Set up logging 13 | logging.basicConfig(level=logging.INFO) 14 | 15 | def load_config(): 16 | """Load configuration settings.""" 17 | return { 18 | 'dynamic_data_path': '.../dynamic_data.csv', 19 | 'static_data_path': '.../static_data.csv', 20 | 'grid_data_path': '.../grid_df.csv', 21 | 'sequence_length': 24, 22 | 'train_start_date': '2019-01-01', 23 | 'train_end_date': '2019-12-31', 24 | 'val_start_date': '2020-01-01', 25 | 'val_end_date': '2020-06-30', 26 | 'test_start_date': '2020-07-01', 27 | 'output_dir': '.../outputs/plots' 28 | } 29 | 30 | 31 | def save_processed_data(output_dir, **data): 32 | """Save processed data to files.""" 33 | os.makedirs(output_dir, exist_ok=True) 34 | 35 | # Save tensors 36 | for key, value in data.items(): 37 | if isinstance(value, torch.Tensor): 38 | file_path = os.path.join(output_dir, f"{key}.pt") 39 | torch.save(value, file_path) 40 | print(f"Saved {key} to {file_path}") 41 | elif isinstance(value, RobustScaler): 42 | file_path = os.path.join(output_dir, f"{key}.pkl") 43 | with open(file_path, 'wb') as f: 44 | import pickle 45 | pickle.dump(value, f) 46 | print(f"Saved {key} to {file_path}") 47 | 48 | 49 | def load_and_prepare_data(config): 50 | """Load and prepare datasets.""" 51 | dynamic_data = pd.read_csv(config['dynamic_data_path']) 52 | static_data = pd.read_csv(config['static_data_path']) 53 | grid_df = pd.read_csv(config['grid_data_path']) 54 | 55 | # Convert datetime and sort data 56 | dynamic_data['datetime'] = pd.to_datetime(dynamic_data['datetime']) 57 | dynamic_data = dynamic_data.sort_values(by=['state', 'datetime']) 58 | dynamic_data['target_consumption'] = dynamic_data.groupby('state')['value'].shift(-1) 59 | dynamic_data = dynamic_data.dropna(subset=['target_consumption']) 60 | 61 | return dynamic_data, static_data, grid_df 62 | 63 | def split_data(dynamic_data, config): 64 | """Split data into train, validation, and test sets.""" 65 | train_data = dynamic_data[(dynamic_data['datetime'] >= config['train_start_date']) & (dynamic_data['datetime'] < config['train_end_date'])].copy() 66 | val_data = dynamic_data[(dynamic_data['datetime'] >= config['val_start_date']) & (dynamic_data['datetime'] < config['val_end_date'])].copy() 67 | test_data = dynamic_data[dynamic_data['datetime'] >= config['test_start_date']].copy() 68 | 69 | for data in [train_data, val_data, test_data]: 70 | data['time_index'] = (data['datetime'] - data['datetime'].min()).dt.total_seconds() // 3600 71 | data['time_index'] = data['time_index'].astype(int) 72 | 73 | return train_data, val_data, test_data 74 | 75 | def scale_features(train_data, val_data, test_data, features_to_scale): 76 | """Scale features using RobustScaler.""" 77 | feature_scaler = RobustScaler() 78 | train_data[features_to_scale] = feature_scaler.fit_transform(train_data[features_to_scale]) 79 | val_data[features_to_scale] = feature_scaler.transform(val_data[features_to_scale]) 80 | test_data[features_to_scale] = feature_scaler.transform(test_data[features_to_scale]) 81 | 82 | return feature_scaler 83 | 84 | def scale_targets(train_data, val_data, test_data): 85 | """Scale target consumption using RobustScaler.""" 86 | target_scaler = RobustScaler() 87 | train_data['target_consumption'] = target_scaler.fit_transform(train_data[['target_consumption']]) 88 | val_data['target_consumption'] = target_scaler.transform(val_data[['target_consumption']]) 89 | test_data['target_consumption'] = target_scaler.transform(test_data[['target_consumption']]) 90 | return target_scaler 91 | 92 | def create_graph(dynamic_data, static_data, grid_df): 93 | """Create directed graph using NetworkX.""" 94 | G = nx.DiGraph() 95 | 96 | # Add states as nodes with attributes from static_data 97 | for index, row in static_data.iterrows(): 98 | G.add_node(row['state'], **row.to_dict()) 99 | 100 | # Add time series data to each node 101 | for state, group in dynamic_data.groupby('state'): 102 | G.nodes[state]['time_series'] = group.to_dict(orient='records') 103 | 104 | # Add edges with attributes from grid_df 105 | for index, row in grid_df.iterrows(): 106 | G.add_edge(row['Source'], row['Target'], **row.to_dict()) 107 | 108 | return G 109 | 110 | def extract_graph_features(G, node_mapping): 111 | """Extract and scale node and edge features.""" 112 | # Extract node features 113 | node_features_list = [] 114 | for node, data in G.nodes(data=True): 115 | features = [data.get(feature) for feature in ['x', 'y', 'pv_pot', 'onw_pot', 'ofw_pot']] 116 | node_features_list.append(features) 117 | 118 | node_features_df = pd.DataFrame(node_features_list, columns=['x', 'y', 'pv_pot', 'onw_pot', 'ofw_pot']) 119 | 120 | # Normalize node features 121 | scaler = RobustScaler() 122 | normalized_node_features = scaler.fit_transform(node_features_df) 123 | node_features_tensor = torch.tensor(normalized_node_features, dtype=torch.float) 124 | 125 | # Extract edge index and edge attributes 126 | edge_index_list = [] 127 | edge_attr_list = [] 128 | for source, target, data in G.edges(data=True): 129 | edge_index_list.append([node_mapping[source], node_mapping[target]]) 130 | edge_attr = [data.get(attr) for attr in ['capacity', 'line_eff', 'line_len', 'line_carrier']] 131 | edge_attr_list.append(edge_attr) 132 | 133 | edge_attr_df = pd.DataFrame(edge_attr_list, columns=['capacity', 'line_eff', 'line_len', 'line_carrier']) 134 | 135 | # Normalize edge attributes 136 | edge_scaler = RobustScaler() 137 | scaled_edge_attrs = edge_scaler.fit_transform(edge_attr_df) 138 | edge_attr_tensor = torch.tensor(scaled_edge_attrs, dtype=torch.float) 139 | 140 | # Convert edge index to tensor 141 | edge_index_tensor = torch.tensor(edge_index_list, dtype=torch.long).t().contiguous() 142 | 143 | return node_features_tensor, edge_index_tensor, edge_attr_tensor 144 | 145 | def create_sequences(data, sequence_length, features_to_scale, node_mapping): 146 | """Create input sequences, targets, and node indices.""" 147 | sequences, targets, nodes = [], [], [] 148 | grouped = data.groupby('state') 149 | 150 | for state, group in grouped: 151 | group = group.sort_values(by='datetime') 152 | values = group[features_to_scale].values 153 | target_values = group['target_consumption'].values 154 | state_idx = node_mapping[state] 155 | 156 | if len(group) >= sequence_length + 1: 157 | for i in range(len(group) - sequence_length): 158 | seq = values[i:i + sequence_length] 159 | tgt = target_values[i + sequence_length] 160 | sequences.append(seq) 161 | targets.append(tgt) 162 | nodes.append(state_idx) 163 | 164 | sequences_tensor = torch.tensor(np.array(sequences), dtype=torch.float) 165 | targets_tensor = torch.tensor(np.array(targets), dtype=torch.float).unsqueeze(-1) 166 | nodes_tensor = torch.tensor(np.array(nodes), dtype=torch.long) 167 | 168 | return sequences_tensor, targets_tensor, nodes_tensor 169 | 170 | def visualize_graph(G, output_dir): 171 | """Visualize and save the graph.""" 172 | os.makedirs(output_dir, exist_ok=True) 173 | 174 | plt.figure(figsize=(10, 8)) 175 | pos = {node: (data.get('x', 0), data.get('y', 0)) for node, data in G.nodes(data=True)} 176 | labels = {node: node for node in G.nodes()} 177 | nx.draw( 178 | G, 179 | pos, 180 | with_labels=True, 181 | labels=labels, 182 | node_color='skyblue', 183 | node_size=500, 184 | edge_color='k', 185 | linewidths=1, 186 | font_size=15, 187 | arrows=True, 188 | ) 189 | plt.title("Directed Network Graph of States and Grid Lines") 190 | 191 | graph_path = os.path.join(output_dir, "network_graph.png") 192 | print(f"Saving graph to {graph_path}") 193 | plt.savefig(graph_path, format='png', dpi=300) 194 | assert os.path.isfile(graph_path), f"Graph not saved at {graph_path}" 195 | plt.close() 196 | print(f"Graph successfully saved to {graph_path}") 197 | 198 | 199 | 200 | # Main preprocessing function 201 | def preprocess_data(config): 202 | dynamic_data, static_data, grid_df = load_and_prepare_data(config) 203 | train_data, val_data, test_data = split_data(dynamic_data, config) 204 | 205 | features_to_scale = ['value', 'pv', 'onw', 'ofw', 206 | 'TOTAL HOURLY RAIN (mm)(mean)', 'TOTAL HOURLY RAIN (mm)(std)', 207 | 'ATMOSPHERIC PRESSURE AT STATION LEVEL (mB)(mean)', 208 | 'ATMOSPHERIC PRESSURE AT STATION LEVEL (mB)(std)', 209 | 'GLOBAL RADIATION (KJ/m²)(mean)', 'GLOBAL RADIATION (KJ/m²)(std)', 210 | 'AIR TEMPERATURE (°C)(mean)', 'AIR TEMPERATURE (°C)(std)', 211 | 'DEW POINT TEMPERATURE (°C)(mean)', 'DEW POINT TEMPERATURE (°C)(std)', 212 | 'REL HUMIDITY FOR THE LAST HOUR (%)(mean)', 213 | 'REL HUMIDITY FOR THE LAST HOUR (%)(std)', 'WIND DIRECTION (gr)(mean)', 214 | 'WIND DIRECTION (gr)(std)', 'WIND MAXIMUM GUST (m/s)(mean)', 215 | 'WIND MAXIMUM GUST (m/s)(std)', 'WIND SPEED (m/s)(mean)', 216 | 'WIND SPEED (m/s)(std)', 'year', 'month', 'day', 'hour', 'dayofweek', 217 | 'weekofyear', 'quarter', 'is_holiday', 'season', 'state_id', 218 | 'total_plant_capacity', 'population', 'GDP'] # Add your features list here 219 | 220 | feature_scaler = scale_features(train_data, val_data, test_data, features_to_scale) 221 | target_scaler = scale_targets(train_data, val_data, test_data) 222 | 223 | G = create_graph(dynamic_data, static_data, grid_df) 224 | node_mapping = {node: idx for idx, node in enumerate(sorted(G.nodes))} 225 | node_to_state = {idx: node for node, idx in node_mapping.items()} 226 | node_features_tensor, edge_index_tensor, edge_attr_tensor = extract_graph_features(G, node_mapping) 227 | 228 | train_seq, train_tgt, train_nodes = create_sequences(train_data, config['sequence_length'], features_to_scale, node_mapping) 229 | val_seq, val_tgt, val_nodes = create_sequences(val_data, config['sequence_length'], features_to_scale, node_mapping) 230 | test_seq, test_tgt, test_nodes = create_sequences(test_data, config['sequence_length'], features_to_scale, node_mapping) 231 | 232 | # Visualize and save the graph 233 | visualize_graph(G, config['output_dir']) 234 | 235 | # Save processed outputs 236 | save_processed_data( 237 | config['output_dir'], 238 | train_seq=train_seq, 239 | train_tgt=train_tgt, 240 | train_nodes=train_nodes, 241 | val_seq=val_seq, 242 | val_tgt=val_tgt, 243 | val_nodes=val_nodes, 244 | test_seq=test_seq, 245 | test_tgt=test_tgt, 246 | test_nodes=test_nodes, 247 | node_features_tensor=node_features_tensor, 248 | edge_index_tensor=edge_index_tensor, 249 | edge_attr_tensor=edge_attr_tensor, 250 | target_scaler=target_scaler 251 | ) 252 | 253 | return train_seq, train_tgt, train_nodes, val_seq, val_tgt, val_nodes, test_seq, test_tgt, test_nodes, node_features_tensor, edge_index_tensor, edge_attr_tensor, target_scaler 254 | 255 | -------------------------------------------------------------------------------- /gat_lstm_model.py: -------------------------------------------------------------------------------- 1 | # gat_lstm_model_early_fusion.py 2 | import torch 3 | import torch.nn as nn 4 | from torch_geometric.nn import GATConv 5 | 6 | class GAT_LSTM(nn.Module): 7 | def __init__(self, node_feature_dim, sequence_feature_dim, gat_out_channels, gat_heads, lstm_hidden_dim, lstm_layers, edge_dim): 8 | super(GAT_LSTM, self).__init__() 9 | 10 | # Edge attribute conditioning 11 | self.edge_attr_transform = nn.Linear(edge_dim, gat_out_channels) 12 | 13 | # Parallel GAT layers 14 | self.gat_1hop_1 = GATConv(node_feature_dim, gat_out_channels, heads=gat_heads, concat=False, edge_dim=gat_out_channels) 15 | self.gat_1hop_2 = GATConv(node_feature_dim, gat_out_channels, heads=gat_heads, concat=False, edge_dim=gat_out_channels) 16 | 17 | self.gat_dropout = nn.Dropout(0.2) 18 | 19 | # The input to LSTM is the sequence feature dimension + the output of the GAT layers 20 | combined_input_dim = sequence_feature_dim + 2 * gat_out_channels 21 | self.lstm = nn.LSTM(combined_input_dim, lstm_hidden_dim, num_layers=lstm_layers, batch_first=True) 22 | self.lstm_dropout = nn.Dropout(0.3) 23 | 24 | # Fully connected layer for final prediction 25 | self.fc = nn.Linear(lstm_hidden_dim, 1) # Output layer to predict target consumption 26 | 27 | def forward(self, sequences, edge_index, edge_attr, node_features, node_indices): 28 | # Edge attribute conditioning 29 | transformed_edge_attr = self.edge_attr_transform(edge_attr) 30 | 31 | # GAT layer 1-hop_1 neighbors 32 | gat_1hop_out_1 = self.gat_1hop_1(node_features, edge_index, transformed_edge_attr) 33 | gat_1hop_out_1 = self.gat_dropout(gat_1hop_out_1) 34 | 35 | # GAT layer 1-hop_2 neighbors 36 | gat_1hop_out_2 = self.gat_1hop_2(node_features, edge_index, transformed_edge_attr) 37 | gat_1hop_out_2 = self.gat_dropout(gat_1hop_out_2) 38 | 39 | # Gather GAT output for each sequence's node index (both 1-hop and 2-hop) 40 | gat_1hop_out_1 = gat_1hop_out_1[node_indices] 41 | gat_1hop_out_2 = gat_1hop_out_2[node_indices] 42 | 43 | # Concatenate the GAT outputs (1-hop_1 and 1-hop_2) along the feature dimension 44 | gat_combined_out = torch.cat((gat_1hop_out_1, gat_1hop_out_2), dim=-1) # [batch_size, 2 * gat_out_channels] 45 | 46 | # Expand the GAT output to match the sequence length and concatenate with the sequence data 47 | gat_combined_out = gat_combined_out.unsqueeze(1).repeat(1, sequences.size(1), 1) # Repeat for each time step 48 | combined_input = torch.cat((sequences, gat_combined_out), dim=-1) # [batch_size, seq_len, seq_feat_dim + 2 * gat_out_channels] 49 | 50 | # LSTM layer processes the combined sequence and GAT data 51 | lstm_out, _ = self.lstm(combined_input) 52 | lstm_out = self.lstm_dropout(lstm_out) 53 | 54 | # Take the last output of LSTM 55 | lstm_out = lstm_out[:, -1, :] # [batch_size, lstm_hidden_dim] 56 | 57 | # Fully connected layer to predict the next hour's consumption 58 | out = self.fc(lstm_out) 59 | 60 | return out 61 | -------------------------------------------------------------------------------- /testing.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import numpy as np 4 | import pandas as pd 5 | import logging 6 | import matplotlib.pyplot as plt 7 | from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score 8 | from scipy.stats import pearsonr 9 | from torch.utils.data import DataLoader, TensorDataset 10 | from gat_lstm_model import GAT_LSTM 11 | from data_preprocessing import preprocess_data, load_config 12 | 13 | # Set up logging 14 | logging.basicConfig(level=logging.INFO) 15 | 16 | def evaluate_model(model, test_loader, target_scaler, node_to_state, test_time_indices, output_dir="outputs"): 17 | """Evaluate the model and save results.""" 18 | model.eval() 19 | test_predictions, test_targets, test_nodes, test_hours = [], [], [], [] 20 | 21 | os.makedirs(output_dir, exist_ok=True) 22 | predictions_path = os.path.join(output_dir, "predictions.csv") 23 | evaluation_metrics_path = os.path.join(output_dir, "evaluation_metrics.txt") 24 | 25 | with torch.no_grad(): 26 | for i, (sequences, targets, nodes) in enumerate(test_loader): 27 | sequences, targets, nodes = sequences.to(device), targets.to(device), nodes.to(device) 28 | node_features = node_features_tensor[nodes] 29 | output = model(sequences, edge_index_tensor, edge_attr_tensor, node_features, nodes) 30 | test_predictions.append(output.cpu().numpy()) 31 | test_targets.append(targets.cpu().numpy()) 32 | test_nodes.append(nodes.cpu().numpy()) 33 | 34 | # Extract time indices (hour of the day) based on batch index 35 | batch_size = targets.shape[0] 36 | batch_time_indices = test_time_indices[i * batch_size:(i + 1) * batch_size] 37 | test_hours.extend(batch_time_indices % 24) # Extract hour (0-23) 38 | 39 | # Concatenate results 40 | test_predictions = np.concatenate(test_predictions).squeeze() 41 | test_targets = np.concatenate(test_targets).squeeze() 42 | test_nodes = np.concatenate(test_nodes).squeeze() 43 | 44 | # Inverse transform predictions and targets to original scale 45 | test_predictions = target_scaler.inverse_transform(test_predictions.reshape(-1, 1)).flatten() 46 | test_targets = target_scaler.inverse_transform(test_targets.reshape(-1, 1)).flatten() 47 | 48 | # Save predictions, actuals, node names, and hours 49 | results_df = pd.DataFrame({ 50 | "Node": [node_to_state.get(node, f"Node {node}") for node in test_nodes], 51 | "Hour": test_hours, 52 | "Actual": test_targets, 53 | "Predicted": test_predictions 54 | }) 55 | results_df.to_csv(predictions_path, index=False) 56 | logging.info(f"Saved predictions and actuals to {predictions_path}") 57 | 58 | # Calculate evaluation metrics 59 | mae = mean_absolute_error(test_targets, test_predictions) 60 | rmse = np.sqrt(mean_squared_error(test_targets, test_predictions)) 61 | mape = np.mean(np.abs((test_targets - test_predictions) / np.clip(test_targets, a_min=1e-8, a_max=None))) * 100 62 | r2 = r2_score(test_targets, test_predictions) 63 | corr_coef, _ = pearsonr(test_targets, test_predictions) 64 | 65 | # Save evaluation metrics 66 | with open(evaluation_metrics_path, "w") as f: 67 | f.write(f"Mean Absolute Error (MAE): {mae}\n") 68 | f.write(f"Root Mean Squared Error (RMSE): {rmse}\n") 69 | f.write(f"Mean Absolute Percentage Error (MAPE): {mape}%\n") 70 | f.write(f"R-squared (R2): {r2}\n") 71 | f.write(f"Pearson Correlation Coefficient: {corr_coef}\n") 72 | logging.info(f"Saved evaluation metrics to {evaluation_metrics_path}") 73 | 74 | # Plot actual vs predicted values over 24-hour period for all nodes 75 | plt.figure(figsize=(12, 6)) 76 | all_actuals, all_predicteds = [], [] 77 | unique_nodes = np.unique(test_nodes) 78 | 79 | for node in unique_nodes: 80 | node_indices = np.where(test_nodes == node)[0][:24] # First 24-hour period 81 | all_actuals.append(test_targets[node_indices]) 82 | all_predicteds.append(test_predictions[node_indices]) 83 | 84 | mean_actual = np.mean(np.array(all_actuals), axis=0) 85 | mean_predicted = np.mean(np.array(all_predicteds), axis=0) 86 | 87 | plt.plot(mean_actual, label='Mean Actual', alpha=0.8) 88 | plt.plot(mean_predicted, label='Mean Predicted', alpha=0.8) 89 | plt.title('Mean Actual vs Predicted Values Over 24-Hour Period (All Nodes)') 90 | plt.xlabel('Hour') 91 | plt.ylabel('Load') 92 | plt.legend() 93 | plt.grid() 94 | plt.savefig(os.path.join(output_dir, "actual_vs_predicted_24hr_all_nodes.png"), dpi=300) 95 | plt.close() 96 | logging.info(f"Saved Actual vs Predicted plot (all nodes) to {os.path.join(output_dir, 'actual_vs_predicted_24hr_all_nodes.png')}") 97 | 98 | # Plot actual vs predicted for each individual node over 24-hour period 99 | plt.figure(figsize=(30, 25)) 100 | rows, cols = 5, 6 # Adjust grid size based on the number of nodes 101 | for i, node in enumerate(unique_nodes[:rows * cols]): 102 | node_indices = np.where(test_nodes == node)[0][:24] 103 | actual = test_targets[node_indices] 104 | predicted = test_predictions[node_indices] 105 | 106 | plt.subplot(rows, cols, i + 1) 107 | plt.plot(actual, label='Actual') 108 | plt.plot(predicted, label='Predicted') 109 | node_name = node_to_state.get(node, f"Node {node}") # Use the actual node name 110 | plt.title(f"Actual vs Predicted - {node_name}") 111 | plt.xlabel('Hour') 112 | plt.ylabel('Load') 113 | plt.legend() 114 | plt.grid() 115 | 116 | plt.tight_layout() 117 | plt.savefig(os.path.join(output_dir, "actual_vs_predicted_24hr_individual_nodes.png"), dpi=300) 118 | plt.close() 119 | logging.info(f"Saved Actual vs Predicted plot (individual nodes) to {os.path.join(output_dir, 'actual_vs_predicted_24hr_individual_nodes.png')}") 120 | 121 | return test_targets, test_predictions, mae, rmse, mape, r2, corr_coef 122 | 123 | # Main function to run testing 124 | if __name__ == "__main__": 125 | # Load configuration and preprocess data 126 | config = load_config() 127 | train_seq, train_tgt, train_nodes, val_seq, val_tgt, val_nodes, test_seq, test_tgt, test_nodes, node_features_tensor, edge_index_tensor, edge_attr_tensor, target_scaler = preprocess_data(config) 128 | 129 | # Extract time indices for test dataset 130 | test_time_indices = np.arange(len(test_tgt)) # Replace this with actual time indices if available 131 | 132 | # Load the model 133 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 134 | model = GAT_LSTM( 135 | node_feature_dim=node_features_tensor.shape[1], 136 | sequence_feature_dim=test_seq.shape[2], 137 | gat_out_channels=64, 138 | gat_heads=8, 139 | lstm_hidden_dim=128, 140 | lstm_layers=4, 141 | edge_dim=edge_attr_tensor.shape[1] 142 | ).to(device) 143 | 144 | model_path = os.path.join(config['output_dir'], "gat_lstm_model.pth") 145 | model.load_state_dict(torch.load(model_path, weights_only=True)) 146 | model.eval() 147 | logging.info(f"Loaded model from {model_path}") 148 | 149 | # Move tensors to device 150 | test_seq, test_tgt, test_nodes = test_seq.to(device), test_tgt.to(device), test_nodes.to(device) 151 | node_features_tensor = node_features_tensor.to(device) 152 | edge_index_tensor = edge_index_tensor.to(device) 153 | edge_attr_tensor = edge_attr_tensor.to(device) 154 | 155 | # Prepare DataLoader 156 | batch_size = 27 157 | test_dataset = TensorDataset(test_seq, test_tgt, test_nodes) 158 | test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False) 159 | 160 | # Create node-to-state mapping 161 | node_mapping = {node: idx for idx, node in enumerate(sorted(np.unique(test_nodes.cpu())))} 162 | node_to_state = {idx: node for node, idx in node_mapping.items()} 163 | 164 | # Evaluate the model 165 | output_dir = config['output_dir'] # Directory to save outputs 166 | test_targets, test_predictions, mae, rmse, mape, r2, corr_coef = evaluate_model( 167 | model=model, 168 | test_loader=test_loader, 169 | target_scaler=target_scaler, 170 | node_to_state=node_to_state, 171 | test_time_indices=test_time_indices, 172 | output_dir=output_dir 173 | ) 174 | 175 | logging.info("Testing completed successfully.") 176 | -------------------------------------------------------------------------------- /training.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import logging 4 | from torch.utils.data import DataLoader, TensorDataset 5 | from torch.optim import Adam 6 | from torch.optim.lr_scheduler import ReduceLROnPlateau 7 | import matplotlib.pyplot as plt 8 | from gat_lstm_model import GAT_LSTM 9 | from data_preprocessing import preprocess_data, load_config 10 | 11 | # Set up logging 12 | logging.basicConfig(level=logging.INFO) 13 | 14 | def train_model(model, train_loader, val_loader, optimizer, scheduler, criterion, num_epochs=200, patience=10, output_dir="outputs"): 15 | """Train the model and save the results.""" 16 | train_losses, val_losses = [], [] 17 | best_val_loss = float('inf') 18 | trigger_times = 0 19 | 20 | os.makedirs(output_dir, exist_ok=True) 21 | model_save_path = os.path.join(output_dir, "gat_lstm_model.pth") 22 | training_log_path = os.path.join(output_dir, "training_log.txt") 23 | 24 | # Open a log file to write training progress 25 | with open(training_log_path, 'w') as log_file: 26 | log_file.write("Epoch, Train Loss, Validation Loss\n") 27 | 28 | for epoch in range(num_epochs): 29 | model.train() 30 | train_loss = 0 31 | for sequences, targets, nodes in train_loader: 32 | sequences, targets, nodes = sequences.to(device), targets.to(device), nodes.to(device) 33 | optimizer.zero_grad() 34 | node_features = node_features_tensor[nodes] 35 | output = model(sequences, edge_index_tensor, edge_attr_tensor, node_features, nodes) 36 | loss = criterion(output, targets) 37 | loss.backward() 38 | optimizer.step() 39 | train_loss += loss.item() 40 | 41 | train_loss /= len(train_loader) 42 | train_losses.append(train_loss) 43 | 44 | model.eval() 45 | val_loss = 0 46 | with torch.no_grad(): 47 | for sequences, targets, nodes in val_loader: 48 | sequences, targets, nodes = sequences.to(device), targets.to(device), nodes.to(device) 49 | node_features = node_features_tensor[nodes] 50 | output = model(sequences, edge_index_tensor, edge_attr_tensor, node_features, nodes) 51 | loss = criterion(output, targets) 52 | val_loss += loss.item() 53 | 54 | val_loss /= len(val_loader) 55 | val_losses.append(val_loss) 56 | 57 | # Log results for this epoch 58 | log_msg = f"Epoch {epoch+1}, Train Loss: {train_loss}, Validation Loss: {val_loss}" 59 | logging.info(log_msg) 60 | log_file.write(f"{epoch+1},{train_loss},{val_loss}\n") 61 | 62 | # Adjust learning rate 63 | scheduler.step(val_loss) 64 | 65 | # Save the best model 66 | if val_loss < best_val_loss: 67 | best_val_loss = val_loss 68 | torch.save(model.state_dict(), model_save_path) 69 | logging.info(f"Saved best model at {model_save_path}") 70 | trigger_times = 0 71 | else: 72 | trigger_times += 1 73 | if trigger_times >= patience: 74 | logging.info("Early stopping triggered!") 75 | break 76 | 77 | # Save loss curves 78 | plt.figure(figsize=(10, 6)) 79 | plt.plot(train_losses, label='Train Loss') 80 | plt.plot(val_losses, label='Validation Loss') 81 | plt.xlabel('Epoch') 82 | plt.ylabel('Loss') 83 | plt.title('GAT-LSTM Training and Validation Loss') 84 | plt.legend() 85 | plt.grid() 86 | loss_curve_path = os.path.join(output_dir, "loss_curve.png") 87 | plt.savefig(loss_curve_path, dpi=300) 88 | plt.close() 89 | logging.info(f"Loss curve saved at {loss_curve_path}") 90 | 91 | return train_losses, val_losses 92 | 93 | # Main function to run training 94 | if __name__ == "__main__": 95 | # Load configuration and preprocess data 96 | config = load_config() 97 | train_seq, train_tgt, train_nodes, val_seq, val_tgt, val_nodes, test_seq, test_tgt, test_nodes, node_features_tensor, edge_index_tensor, edge_attr_tensor, target_scaler = preprocess_data(config) 98 | 99 | # Move tensors to device 100 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 101 | train_seq, train_tgt, train_nodes = train_seq.to(device), train_tgt.to(device), train_nodes.to(device) 102 | val_seq, val_tgt, val_nodes = val_seq.to(device), val_tgt.to(device), val_nodes.to(device) 103 | node_features_tensor = node_features_tensor.to(device) 104 | edge_index_tensor = edge_index_tensor.to(device) 105 | edge_attr_tensor = edge_attr_tensor.to(device) 106 | 107 | # Prepare DataLoader 108 | batch_size = 27 109 | train_dataset = TensorDataset(train_seq, train_tgt, train_nodes) 110 | val_dataset = TensorDataset(val_seq, val_tgt, val_nodes) 111 | train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=False) 112 | val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False) 113 | 114 | # Model parameters 115 | sequence_feature_dim = train_seq.shape[2] 116 | node_feature_dim = node_features_tensor.shape[1] 117 | gat_out_channels = 64 118 | gat_heads = 8 119 | lstm_hidden_dim = 128 120 | lstm_layers = 4 121 | edge_dim = edge_attr_tensor.shape[1] 122 | 123 | # Initialize model 124 | model = GAT_LSTM(node_feature_dim, sequence_feature_dim, gat_out_channels, gat_heads, lstm_hidden_dim, lstm_layers, edge_dim).to(device) 125 | 126 | # Define optimizer, scheduler, and loss function 127 | optimizer = Adam(model.parameters(), lr=0.0001, weight_decay=1e-5) 128 | scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=5, verbose=True) 129 | criterion = torch.nn.MSELoss() 130 | 131 | # Train the model 132 | output_dir = config['output_dir'] # Directory to save outputs 133 | train_losses, val_losses = train_model( 134 | model=model, 135 | train_loader=train_loader, 136 | val_loader=val_loader, 137 | optimizer=optimizer, 138 | scheduler=scheduler, 139 | criterion=criterion, 140 | num_epochs=200, 141 | patience=10, 142 | output_dir=output_dir 143 | ) 144 | 145 | logging.info("Training completed successfully.") 146 | --------------------------------------------------------------------------------