Model training methods, power system load forecasting methods, devices and electronic equipment

Abstract

This application discloses a model training method and a power system load forecasting method. The model training method uses a recurrent neural network as the architecture of the policy model. The model parameters of the policy model are iteratively updated through the GRPO algorithm. The GRPO algorithm can provide more intelligent feature selection and task allocation, enabling the policy model to adaptively select different policies to participate in the forecast, thereby improving the model's generalization ability and dynamic modeling ability. Then, the simulated annealing algorithm is used to optimize the hyperparameters of the policy model, such as network layers, number of neurons, and learning rate, so that the model can escape local optima and improve the global prediction performance of the model in multimodal feature environments. The training method proposed in this invention combines the GRPO algorithm, recurrent neural network, and simulated annealing algorithm, which can effectively improve the accuracy of load forecasting, enhance the model's adaptability to multi-source inputs, and exhibit better stability and scalability in power system scenarios with strong data nonlinearity, multi-scale, and multi-task characteristics.

Description

Technical Field

[0001] This invention pertains to power system load forecasting technology, and particularly relates to a model training method, a power system load forecasting method, a device, and electronic equipment. Background Technology

[0002] Early load forecasting was mainly based on traditional statistical methods, such as moving average, autoregressive moving average (ARMA) model, and seasonal autoregressive integral moving average (SARIMA) model. These methods performed well when the load data characteristics were relatively stable. However, with the increasing complexity of power systems and the enhanced nonlinearity of load data, traditional methods have gradually become unable to meet modern needs.

[0003] With the application of artificial intelligence in various fields, load forecasting has gradually integrated machine learning and artificial intelligence technologies. Machine learning methods, such as Support Vector Machines (SVM) and Random Forests (RF), have provided new approaches for modeling complex load change patterns. Subsequently, deep learning models (such as Long Short-Term Memory Networks (LSTM) and Convolutional Neural Networks (CNN)) have rapidly become the mainstream technology in the field of load forecasting due to their superior performance in processing large-scale nonlinear data and mining complex features. Supported by big data and computing resources, these methods can significantly improve the accuracy of load forecasting.

[0004] However, the power system load is affected by various factors such as temperature, humidity, time period, and economic activities, resulting in fluctuations. The load data has complex nonlinear and random characteristics, which makes it easy for the optimization problem of load forecasting to get trapped in local optima. In addition, the power load has complex characteristics such as periodicity, seasonality, and randomness, and the existing forecasting models still do not have good generalization ability. Summary of the Invention

[0005] Based on this, the present invention aims to propose a model training method, a power system load forecasting method, a device and electronic equipment, to construct a load forecasting strategy model based on LSTM network, to train and optimize the model using the GRPO algorithm, and to optimize the hyperparameters of the model using the simulated annealing algorithm, so as to enhance the model's adaptability to complex multi-source inputs and improve the global forecasting performance in a multi-task environment.

[0006] In a first aspect, the present invention provides a model training method, wherein the trained model is used for power system load forecasting, comprising:

[0007] Historical power system load data was obtained as a training set.

[0008] A strategy model for load forecasting is established based on a recurrent neural network, and the training set is used as the input to the strategy model for model training.

[0009] The GRPO algorithm is used to iteratively update the model parameters of the strategy model, and the simulated annealing algorithm is used to update the hyperparameters of the strategy model. The trained strategy model is then used as the load prediction model.

[0010] Furthermore, before training the model by using the training set as input to the policy model, the following steps are included:

[0011] Construct a global hint vector;

[0012] The global cue vector is concatenated with the training set, and the training set is updated to the concatenated training set, which is then used as the input to the policy model.

[0013] Furthermore, the model parameters of the strategy model iteratively updated using the GRPO algorithm include:

[0014] Calculate policy loss based on the GRPO algorithm;

[0015] Calculate the vector loss of the global cue vector based on the policy loss;

[0016] An optimization objective function for the policy model is constructed based on policy loss and vector loss. The optimization objective function is iteratively solved to update the parameters of the policy model.

[0017] Furthermore, the strategy loss is calculated as follows:

[0018] ,

[0019] in, This indicates the calculation of the expected value; This represents the policy loss calculated based on the GRPO algorithm; Indicates the strategy parameters; Indicates the strategy ratio, Measure the new strategy In state Generate execution action The probability compared to the old strategy The change; clip represents the clipping operation in the GRPO algorithm, used to limit the range of policy variation. , The shear coefficient is... This represents the advantage estimation in the GRPO algorithm.

[0020] Furthermore, the vector loss for calculating the global cue vector based on the policy loss includes:

[0021] The relationship between policy loss and vector loss is established as follows:

[0022] ,

[0023] in, The vector loss represents the global cue vector. The weights represent the vector loss coefficients. Indicates the strategy loss. This represents the descent gradient of the policy loss with respect to the global cue vector p.

[0024] Furthermore, an optimization objective function for the policy model is constructed based on policy loss and vector loss. The optimization objective function is iteratively solved to update the various model parameters of the policy model, including:

[0025] The objective function for constructing the policy model is as follows:

[0026] ,

[0027] in, This represents the total loss of the policy model. Indicates the strategy loss. Represents vector loss. , and Indicates the weighting coefficient. Represents the entropy regularization term. The policy represents the policy model;

[0028] Update the policy parameters and global hint vector based on the following relationship:

[0029]

[0030] ,

[0031] in, Indicates the strategy parameters, Represents the global hint vector. This represents the learning rate of the policy model. Indicates total loss Regarding strategy parameters The descent gradient, Indicates total loss Regarding global hint vectors The descent gradient.

[0032] Furthermore, obtaining historical power system load data as a training set includes:

[0033] Obtain historical data on power system load;

[0034] Feature extraction is performed on historical data to obtain load features for at least one modality;

[0035] The load characteristics of at least one mode include at least one of the following: time characteristics, meteorological characteristics, load history characteristics, economic characteristics, and power system operation characteristics;

[0036] Use the load features of at least one modality as the training set.

[0037] Secondly, this invention proposes a power system load forecasting method, comprising:

[0038] Obtain historical data on power system load;

[0039] Historical data is input into the load forecasting model trained by the first method to obtain the load forecasting results of the power system.

[0040] Thirdly, the present invention proposes a model training device, comprising:

[0041] The training set acquisition module is used to acquire historical data of power system load as a training set.

[0042] The model training module is used to build a strategy model for load forecasting based on a recurrent neural network, and to train the model by using the training set as input to the strategy model.

[0043] The model parameter update module is used to iteratively update the model parameters of the strategy model using the GRPO algorithm and to update the hyperparameters of the strategy model using the simulated annealing algorithm, so as to use the trained strategy model as the load prediction model.

[0044] Fourthly, the present invention provides a power system load forecasting device, comprising:

[0045] The historical data acquisition module is used to acquire historical data on the power system load.

[0046] The load forecasting module is used to input historical data into the load forecasting model trained by a third-party device to obtain the load forecasting results of the power system.

[0047] Fifthly, the present invention provides an electronic device including a memory storing computer-executable instructions and a processor, wherein when the computer-executable instructions are executed by the processor, the device performs various steps of the model training method provided in the first aspect and / or the power system load forecasting method provided in the second aspect.

[0048] In a sixth aspect, the present invention provides a readable storage medium storing a computer-executable program that, when executed, can implement the model training method provided in the first aspect and / or the various steps of the power system load forecasting method provided in the second aspect.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention proposes a model training method and a power system load forecasting method. The model training process uses a recurrent neural network as the architecture for the policy model. The model parameters are iteratively updated using the GRPO algorithm, which provides more intelligent feature selection and task allocation, enabling the policy model to adaptively select different policies for forecasting, thus improving the model's generalization and dynamic modeling capabilities. Then, simulated annealing is used to optimize the hyperparameters of the policy model, such as network layers, number of neurons, and learning rate, allowing the model to escape local optima and improving its global prediction performance in multimodal environments. The training method proposed in this invention combines the GRPO algorithm, recurrent neural network, and simulated annealing algorithm, effectively improving the accuracy of load forecasting, enhancing the model's adaptability to multi-source inputs, and exhibiting better stability and scalability in power system scenarios with strong data nonlinearity, multi-scale, and multi-task characteristics. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the implementation of a model training method provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the simulated annealing algorithm provided in this embodiment of the invention for finding the optimal solution;

[0054] Figure 3 This is a schematic diagram illustrating how the prediction error changes with the number of training rounds in the model training method provided in this embodiment of the invention.

[0055] Figure 4 This is a flowchart illustrating the implementation of a power system load forecasting method provided in an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram of a model training device structure provided in an embodiment of the present invention;

[0057] Figure 6 This is a schematic diagram of the structure of a power system load forecasting device provided in an embodiment of the present invention;

[0058] Figure 7 This is an electronic device architecture diagram provided for an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The GRPO algorithm used in the following embodiments of the present invention is an improvement on the traditional reinforcement learning algorithm PPO algorithm.

[0061] The Proximal Policy Optimization (PPO) algorithm is a widely used reinforcement learning algorithm belonging to the decision-evaluation type. It requires training a value function to estimate the advantage function. The core idea is to avoid excessive policy updates by limiting the magnitude of policy updates. However, the value function network in the PPO algorithm is usually another model of similar size to the policy model, which brings a large amount of memory and computational burden.

[0062] As an improvement on the PPO algorithm, the GRPO algorithm (Group Relative Policy Optimization) removes the value function network and optimizes the policy model through intra-group relative rewards. Unlike PPO, it doesn't require additional value function approximation; instead, it uses the average reward of multiple sampled outputs generated for the same problem as a baseline. Specifically, it samples a set of actions for each state and adjusts the policy based on the relative performance of these actions, rather than relying on a single value network to estimate the value of each action. For each problem q, it samples a set of outputs from the old policy and optimizes the policy model by maximizing its defined objective. The GRPO algorithm leverages the advantages of group-relative computation, which aligns well with the comparative nature of reward models, as reward models are typically trained on datasets comparing outputs for the same problem.

[0063] Based on this, the following embodiments of the present invention will propose a model training method for training a load prediction model, using a recurrent neural network as the architecture of the prediction model, using the GRPO algorithm to optimize the model parameters, and using the simulated annealing algorithm to optimize and update the model's hyperparameters.

[0064] For details, please refer to [link / reference]. Figure 1 An embodiment of the present invention provides a model training method, wherein the trained model is used for power system load forecasting, comprising the following steps:

[0065] Step S110. Obtain historical data of power system load as a training set.

[0066] Specifically, this step collects the actual load values ​​of the power system at various time periods (such as minute, hour, and day). Furthermore, it can also collect weather data (such as temperature, humidity, and sunshine level) and date information (such as holidays and workdays) that are synchronized with the historical load data in the time dimension.

[0067] Furthermore, multimodal characteristics related to load changes are recorded synchronously, including time characteristics, meteorological characteristics, historical load characteristics, economic characteristics, and power system operation characteristics, and the load characteristics of at least one mode are used as the training set.

[0068] In a further embodiment, before the input model is trained, the training set data is preprocessed as necessary, including missing value imputation, denoising, normalization, etc.

[0069] Furthermore, to guide the policy model to more stably perceive long-term dependencies or cross-task differences in data, the model's input training set is modified. Specifically, the initial input is defined as... The preprocessed input is represented as Define a global hint vector As a global status / task prompt, it is concatenated with the input as follows:

[0070]

[0071] Implement the input of the policy model from Transformed into .

[0072] The initial value of the global cue vector p can be generated by a random function and then updated during the iterative update of policy parameters using the GRPO algorithm.

[0073] Step S120. Establish a strategy model for load forecasting based on a recurrent neural network, and train the model by using the training set as input to the strategy model.

[0074] The strategy model in this step uses a recurrent neural network structure. Further implementations can choose LSTM or GRU networks. For the constructed strategy model, the strategy is defined as the output load prediction or load distribution probability for the input state sequence, i.e., historical load data, using weight coefficients representing the strategy model, and iteratively updated using the GRPO algorithm.

[0075] Step S130. Iteratively update the model parameters of the strategy model using the GRPO algorithm, and update the hyperparameters of the strategy model using the simulated annealing algorithm, and use the trained strategy model as the load prediction model.

[0076] Specifically, the GRPO algorithm works by comparing multiple parallel-sampled policies, selecting the best-performing policy sample, and guiding policy improvement.

[0077] Furthermore, for the case of concatenating global suggestion vectors into the model input as described in the aforementioned embodiments, the step of iteratively updating the model parameters of the strategy model using the GRPO algorithm includes the following execution process:

[0078] Step S131. Calculate the policy loss based on the GRPO algorithm.

[0079] The policy loss calculated in this step uses the loss calculation part of the original optimization objective function defined in the GRPO algorithm, which can be expressed as:

[0080]

[0081] in, This indicates the calculation of the expected value; This represents the policy loss calculated based on the GRPO algorithm; Indicates the strategy parameters; Indicates the strategy ratio, Measure the new strategy In state Generate execution action The probability compared to the old strategy The change; clip represents the clipping operation in the GRPO algorithm, used to limit the range of policy variation. , The shear coefficient is... This represents the advantage estimation in the GRPO algorithm.

[0082] Step S132. Calculate the vector loss of the global cue vector based on the policy loss.

[0083] The purpose of this step is to guide the feature learning direction of the global cue vector, ensuring that it contributes positively to policy optimization. Therefore, a regularization term for the global cue vector is constructed to encourage the cue vector to optimize the policy loss on gradient descent, which can be defined as follows:

[0084]

[0085] in, The vector loss represents the global cue vector. The weights represent the vector loss coefficients. Indicates the strategy loss. This represents the descent gradient of the policy loss with respect to the global cue vector p.

[0086] Step S133. Construct an optimization objective function for the policy model based on policy loss and vector loss, and iteratively solve the optimization objective function to calculate the updated values ​​of each policy parameter of the policy model.

[0087] This step considers both policy parameters and global cue vectors simultaneously to minimize the overall loss of the load forecasting task. Therefore, the final optimization objective of the policy model can be expressed as:

[0088]

[0089] in, This represents the total loss of the policy model. Indicates the strategy loss. Represents vector loss. , and Indicates the weighting coefficient. This represents the entropy regularization term.

[0090] The policy parameters and global suggestion vector are jointly optimized, and backpropagation is performed to complete the following updates:

[0091]

[0092]

[0093] in, This represents the learning rate of the policy model. Indicates total loss Regarding strategy parameters The descent gradient, Indicates total loss Regarding global hint vectors The descent gradient.

[0094] Repeat the above parameter update steps until the objective function converges.

[0095] Furthermore, taking the LSTM network as the policy model as an example, the process of updating the model parameters of the GRPO algorithm is explained as follows:

[0096] For each time step, the policy model and LSTM network generate the following actions based on the input:

[0097]

[0098] in, This represents the state at the current time t. This represents the hidden state of the LSTM network at the previous time step t-1, and the action to be performed. And observe the state at the next moment. Strategy reward is defined as ,Bundle Stored in the trajectory buffer.

[0099] Based on the definition of the objective function of the GRPO algorithm, the following policy loss is calculated:

[0100]

[0101] The strategy ratio is expressed as follows:

[0102]

[0103] Advantage estimation Used to measure the state Next action How much better a strategy is than the average can be estimated using the GAE (Generalized Advantage Estimation) method; shearing coefficient. The value is usually set to 0.1 or 0.2 to control the range of updates and prevent the policy updates from being too large, which could lead to training instability.

[0104] Introducing a regularization term from the GRPO algorithm into the overall optimization objective of the policy model, i.e., the regularization objective of the GRPO algorithm on the global cue vector, is expressed as:

[0105]

[0106] The final objective function of the policy model, i.e., the total loss, is:

[0107]

[0108] in, This represents the total loss of the policy model; Indicates the strategy loss; Represents vector loss; , and Indicates the weighting coefficient;

[0109] Representing the entropy regularization term, the policy model for state s can be expressed as an action probability distribution. Then the entropy of each state s can be calculated as follows:

[0110]

[0111] The average entropy of each state is obtained. .

[0112] After calculating the optimization objective, the parameters are iteratively optimized, and the Adam optimizer is used to update the following model parameters:

[0113]

[0114]

[0115] in, This represents the learning rate of the policy model. Indicates total loss Regarding strategy parameters The descent gradient, Indicates total loss Regarding global hint vectors The descent gradient.

[0116] Furthermore, using simulated annealing to update the hyperparameters of the policy model involves updating the configuration parameters that control the training and structure of the network, such as the learning rate, the number of layers in the policy network, the dropout ratio, the selection of input feature dimensions, and the length of the historical window.

[0117] Specifically, the execution process of the simulated annealing algorithm includes:

[0118] (1) Initial solution based on experience or random selection strategy model.

[0119] (2) Generating a neighborhood solution: Based on the current solution, randomly fine-tuning some parameters to generate a new solution. If the current solution is X, then the new solution... This is achieved by fine-tuning parameters, such as slightly increasing the learning rate or changing the number of layers in the policy network.

[0120] (3) Criteria for accepting a new solution: Calculate the objective function value of the new solution. and the target value of the current solution In comparison, if If the new solution is better, then the new solution should be accepted directly. Then there is a certain probability of accepting the new solution. The probability of accepting the new solution is calculated according to the Monte Carlo criterion as follows:

[0121]

[0122] Where T is the current temperature. The higher the temperature, the greater the probability of accepting a worse solution and the stronger the exploration. The lower the temperature, the smaller the probability of acceptance and the more gradually the solution converges.

[0123] (4) Temperature Decrease: After each iteration, the temperature is decreased according to the cooling rate, such as... , This indicates the cooling rate. As the temperature decreases, the randomness of the algorithm gradually decreases, and the algorithm moves from the exploration phase to the convergence phase.

[0124] (5) Iterative update: Stop iterating when the temperature drops to the minimum value or the number of iterations reaches the upper limit, and output the optimal solution, that is, the optimal model hyperparameters.

[0125] Figure 2 This illustrates the process of a simulated return algorithm finding the optimal solution. Figure 3 This illustrates the process of the total loss of the cost prediction model proposed in this invention converging with each training round.

[0126] The above embodiments propose a model training method that uses a recurrent neural network as the architecture of the policy model. The model parameters of the policy model are iteratively updated through the GRPO algorithm. The GRPO algorithm can provide more intelligent feature selection and task allocation, enabling the policy model to adaptively select different policies to participate in prediction, thereby improving the model's generalization ability and dynamic modeling ability. Then, the simulated annealing algorithm is used to optimize the hyperparameters of the policy model, such as network layers, number of neurons, and learning rate, so that the model can escape local optima and improve the global prediction performance of the model in multimodal feature environments. The training method proposed in this invention combines the GRPO algorithm, recurrent neural network, and simulated annealing algorithm, which can effectively improve the accuracy of load forecasting, enhance the model's adaptability to multi-source inputs, and exhibit better stability and scalability in power system scenarios with strong data nonlinearity, multi-scale, and multi-task characteristics.

[0127] See Figure 4 An embodiment of the present invention provides a power system load forecasting method, comprising the following steps:

[0128] Step S410. Obtain historical data on power system load.

[0129] Step S420. Input historical data into the trained load forecasting model to obtain the load forecasting results of the power system.

[0130] The disclosed method can be implemented using various types of devices. Therefore, the present invention also discloses an apparatus corresponding to the above method, and specific embodiments are given below for detailed description.

[0131] like Figure 5 As shown, one embodiment of the present invention provides a model training apparatus for training a load prediction model, comprising:

[0132] The training set acquisition module 502 is used to acquire historical data of power system load as a training set.

[0133] The model training module 504 is used to build a strategy model for load forecasting based on a recurrent neural network, and to train the model by using the training set as input to the strategy model.

[0134] The model parameter update module 506 is used to iteratively update the model parameters of the strategy model using the GRPO algorithm and to update the hyperparameters of the strategy model using the simulated annealing algorithm, so as to use the trained strategy model as the load prediction model.

[0135] See Figure 6 An embodiment of the present invention provides a power system load forecasting device, comprising:

[0136] Historical data acquisition module 602 is used to acquire historical data of power system load;

[0137] The load forecasting module 604 is used to input historical data into the load forecasting model trained by the aforementioned model training device to obtain the load forecasting results of the power system.

[0138] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0139] The methods and related apparatuses mentioned in the above embodiments are described with reference to the method flowcharts and / or structural diagrams provided in the embodiments of this application. Specifically, each block of the method flowchart and / or structural diagram, as well as combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 A process or multiple processes and / or structures illustrate the steps of the functions specified in one or more boxes.

[0140] The following embodiments illustrate the application of this method to a computer device. It is understood that the computer device can be any device with computing and processing capabilities, including but not limited to servers or personal laptops. In one embodiment, the computer device can be an application server, which can be a server used to run the application under test.

[0141] See Figure 7 This document illustrates a hardware block diagram of an electronic device intended to represent various forms of digital computers, such as laptops, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present application described and / or claimed herein.

[0142] like Figure 7 As shown, the electronic device includes: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4;

[0143] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;

[0144] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0145] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0146] The memory stores a program, which the processor can call. The program is used to implement the aforementioned model training method and / or the various steps of the power system load forecasting method.

[0147] This invention also provides a readable storage medium storing a computer program thereon, which, when executed by a processor, implements the model training method and / or the various steps of the power system load forecasting method provided in any possible implementation of the above embodiments and / or in combination with the embodiments.

[0148] The invention has been described in particular detail above with respect to possible scenarios, and those skilled in the art will recognize that the invention can be practiced through other embodiments. Specific naming of components, capitalization of terms, attributes, data structures, or any other programming or structural aspects are not mandatory or important, and the mechanisms or features of implementing the invention may have different names, forms, or procedures. The system can be implemented through a combination of hardware and software (as described), entirely through hardware elements, or entirely through software elements. The specific division of functions among the various system components described herein is merely exemplary and not mandatory; rather, the functions performed by a single system component can be performed by multiple components, or the functions performed by multiple components can be performed by a single component.

[0149] Those skilled in the art should understand that the various steps of the disclosed methods can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, which can then be stored in a storage device for execution by the computing device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, the embodiments disclosed in this invention are not limited to any specific hardware and software combination.

[0150] The programs (also referred to as programs, software, software applications, or code) executable by these computing devices include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0151] Certain aspects of this invention include the process steps and instructions described herein in algorithmic form. It should be noted that the process steps and instructions of this invention can be implemented in software, firmware, and / or hardware, and when implemented in software, they can be downloaded, stored on various operating systems and operated from said platforms.

[0152] Those skilled in the art will understand that the structures shown in the figures are merely block diagrams of some structures related to the present application and do not constitute a limitation on the terminal device to which the present application is applied. Specific terminal devices may include more or fewer components than those shown in the figures, or combine certain components, or have different component arrangements.

[0153] In the description of this specification, the use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "possible design," etc., refers to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0154] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model training method, characterized in that, The trained model is used for power system load forecasting, including: Historical power system load data was obtained as a training set. A strategy model for load forecasting is established based on a recurrent neural network, and the training set is used as the input to the strategy model for model training. The model parameters of the strategy model are iteratively updated using the intra-group relative strategy optimization algorithm, and the hyperparameters of the strategy model are updated using the simulated annealing algorithm. The trained strategy model is then used as the load prediction model. Before using the training set as input to the policy model for model training, the following steps are included: Construct a global hint vector; The global cue vector is concatenated with the training set to obtain a concatenated training set, which is used as the input to the policy model. The model parameters of the strategy model iteratively updated using the within-group relative strategy optimization algorithm include: The policy loss is calculated based on the intra-group relative policy optimization algorithm. Calculate the vector loss of the global cue vector based on the policy loss; An optimization objective function for the policy model is constructed based on the policy loss and vector loss, and the optimization objective function is iteratively solved to update the model parameters of the policy model.

2. The method according to claim 1, characterized in that, The step of calculating the vector loss of the global cue vector based on the policy loss includes: The relationship between the policy loss and the vector loss is established as follows: ， in, The vector loss represents the global cue vector. The weights represent the vector loss coefficients. Indicates the strategy loss. This represents the descent gradient of the policy loss with respect to the global cue vector p.

3. The method according to claim 1, characterized in that, The optimization objective function for constructing the policy model based on the policy loss and vector loss, and the iterative solution of the optimization objective function to update the model parameters of the policy model, include: The objective function for constructing the policy model is as follows: ， in, This represents the total loss of the policy model. Indicates the strategy loss. Represents vector loss. , and Indicates the weighting coefficient. Represents the entropy regularization term. The policy represents the policy model; Update the policy parameters and global hint vector based on the following relationship: ， ， in, Indicates the strategy parameters, Represents the global hint vector. This represents the learning rate of the policy model. Indicates total loss Regarding strategy parameters The descent gradient, Indicates total loss Regarding global hint vectors The descent gradient.

4. A power system load forecasting method, characterized in that, include: Obtain historical data on power system load; The historical data is input into the load forecasting model trained by the model training method described in any one of claims 1 to 3 to obtain the load forecasting results of the power system.

5. A model training device, characterized in that, include: The training set acquisition module is used to acquire historical data of power system load as a training set. The model training module is used to build a strategy model for load forecasting based on a recurrent neural network, and to train the model by using the training set as input to the strategy model. The model parameter update module is used to iteratively update the model parameters of the strategy model using the intra-group relative strategy optimization algorithm, and to update the hyperparameters of the strategy model using the simulated annealing algorithm, so as to use the trained strategy model as the load prediction model. The process of using the training set as input to the policy model for model training includes: Construct a global hint vector; The global cue vector is concatenated with the training set to obtain a concatenated training set, which is used as the input to the policy model. The model parameters of the strategy model iteratively updated using the within-group relative strategy optimization algorithm include: The policy loss is calculated based on the intra-group relative policy optimization algorithm. Calculate the vector loss of the global cue vector based on the policy loss; An optimization objective function for the policy model is constructed based on the policy loss and vector loss, and the optimization objective function is iteratively solved to update the model parameters of the policy model.

6. A power system load forecasting device, characterized in that, include: The historical data acquisition module is used to acquire historical data on the power system load. The load forecasting module is used to input historical data into the load forecasting model trained by the device as described in claim 5 to obtain the load forecasting results of the power system.

7. An electronic device, characterized in that, It includes a memory storing computer-executable instructions and a processor, which, when executed by the processor, causes the device to perform the model training method as described in any one of claims 1 to 3, and / or the power system load forecasting method as described in claim 4.

8. A readable storage medium, characterized in that, It stores a computer-executable program that, when executed, can implement the model training method as described in any one of claims 1 to 3, and / or the power system load forecasting method as described in claim 4.

Images