End-to-end multi-task load prediction method for day-ahead economic dispatching

Abstract

The invention discloses an end-to-end multi-task load prediction method for day-ahead economic dispatching, and the method comprises the following steps: S1, solving an economic dispatching problem, building a neural network training input-output relation, obtaining a learning-optimization model, and storing the learning-optimization model as a pre-training model; s2, designing a deep learning structure of double encoders-single decoder based on a sequence-to-sequence algorithm, respectively learning dependency and features in time sequence data through two encoders, and fusing information into the decoder for unified decoding; and S3, designing a parallel load prediction framework based on the learning-optimization model and the encoder-decoder structure. According to the method, the prediction precision and the scheduling efficiency are remarkably improved by combining the multi-task learning technology and the end-to-end learning technology. According to the scheme, load prediction and economic dispatching tasks are uniformly optimized in the same framework, so that a more efficient and more economic power dispatching strategy is realized.

Description

Technical Field

[0001] The present invention relates to the field of load forecasting and economic dispatch optimization in power systems, and in particular to an end-to-end multi-task load forecasting method for day-ahead economic dispatching. Background Art

[0002] With the transformation of the global energy structure and the rapid development of renewable energy technology, the operating environment of the power system has become more complex. The access of renewable energy, especially wind and solar energy, which are highly random and intermittent, has brought new challenges to traditional power dispatch strategies.

[0003] In the existing technology, although load forecasting and economic dispatch play a vital role in the power system, traditional methods still face many problems. First, traditional load forecasting methods mainly rely on historical data and are modeled through statistical models such as linear regression and SARIMA. However, these methods show obvious limitations when dealing with complex nonlinear relationships and emergencies in the power system, resulting in prediction accuracy that is difficult to meet actual dispatching needs.

[0004] Secondly, although deep learning technologies (such as LSTM, CNN, GCN, etc.) have made some progress in improving load forecasting accuracy, their high hyperparameter sensitivity and computational efficiency issues when dealing with large systems limit their widespread application. In addition, most existing studies separate load forecasting from economic dispatch tasks, ignoring the synergy between the two, resulting in a misalignment between the forecast results and the dispatch objectives in practical applications, and failing to effectively reduce dispatch costs.

[0005] Finally, although end-to-end learning methods have begun to explore the integration of forecasting and scheduling tasks in the same optimization framework, these methods mainly focus on deterministic load forecasting and have poor adaptability to uncertain and complex systems. Therefore, how to organically combine load forecasting accuracy with scheduling task optimization to improve overall system performance remains a major challenge in existing technologies. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an end-to-end multi-task load forecasting method for day-ahead economic dispatch, which significantly improves the forecasting accuracy and dispatching efficiency by combining multi-task learning with end-to-end learning technology. The scheme optimizes load forecasting and economic dispatching tasks in the same framework, thereby achieving a more efficient and economical power dispatching strategy.

[0007] An end-to-end multi-task load forecasting method for day-ahead economic dispatch comprises the following steps:

[0008] S1. According to the standard day-ahead economic dispatch algorithm, the economic dispatch problem is constructed and solved, and then the input-output relationship of the neural network is trained to obtain the learning-optimization model, which is saved as a pre-trained model;

[0009] When traditional optimizers process large-scale data, the training process is often time-consuming and brings a huge computational burden. The present invention greatly accelerates the reasoning speed of the model through the learning-optimization model and the mapping relationship between the input and output of the learning solver. The learning-optimization structure, with its efficient computing performance and low memory usage, can significantly reduce the training time while maintaining the prediction accuracy. In addition, the application of this technology in the multi-task learning framework enables the model to quickly adapt to the prediction tasks of different loads, enhances the generalization ability and real-time response ability of the model, and greatly reduces the computational cost of load forecasting and economic dispatch in large-scale power systems.

[0010] S2. Design a deep learning structure of dual encoders and single decoders based on the sequence-to-sequence algorithm. Use two encoders to learn the dependencies and features in the time series data respectively, and fuse the information into the decoder for unified decoding.

[0011] A custom encoder-decoder structure is used to handle multi-step load forecasting tasks. The encoder part extracts the time series characteristics and spatial characteristics of the input data through a multi-layer neural network structure, which enhances the model's perception of time series relationships and the capture of local features. The decoder part integrates and decodes the encoded information to generate more accurate load forecasting results. Through this structure, the model can effectively improve the accuracy of multi-step load forecasting, greatly improving the accuracy and adaptability of the forecast, especially when dealing with load forecasting in large-scale AC power systems.

[0012] S3. Based on the learning-optimization model and the encoder-decoder structure, a parallel load forecasting framework is designed. Multi-task learning is performed through the parallel encoder-decoder structure. At the same time, through the designed loss function and combined with the learning-optimization model, the accuracy of economic dispatch is considered on the basis of prediction accuracy, making the load forecasting result more suitable for the day-ahead economic dispatch of the power system.

[0013] The customized encoder-decoder structure and the loss function that unifies the load forecast and the dispatching target are designed. This loss function not only focuses on the accuracy of the forecast, but also directly takes the dispatching cost as part of the optimization target. Through this design, the model not only seeks to minimize the forecast error during the training process, but also optimizes the dispatching strategy to achieve the global optimal economic dispatching result. The design of this loss function avoids the problem of poor dispatching effect caused by inconsistent targets in traditional methods, so that the forecast results can better serve the actual dispatching needs.

[0014] In step S3, at the beginning of training, historical load data of a specified time length is selected for each load bus and input into the corresponding dual encoder-single decoder structure for load prediction, and each dual encoder-single decoder structure outputs prediction results of multiple time steps; then the prediction results of multiple time steps are input into the learning-optimization model to obtain the prediction results of the lowest scheduling cost; and on this basis, a loss function is designed, and based on the loss function, the training update of the learning-optimization model and each dual encoder-single decoder structure is realized;

[0015] In this process, a loss function that unifies load forecasting and dispatching objectives is established. This loss function not only focuses on the accuracy of the forecast, but also directly takes the dispatching cost as part of the optimization objective, including:

[0016] For each batch prediction, the output matrix size of each dual encoder-decoder structure is [batch size × 24]. Assuming that there are n different load forecasting tasks in the power system, n = N, each load forecast corresponds to a load bus, and the predicted load is also called the load data of the load bus. The prediction accuracy loss of each batch prediction of each load forecasting task is:

[0017]

[0018] Among them, B represents the batch size, that is, the number of samples in each batch; the number 24 represents the number of time steps, corresponding to 24 hours in a day; is the actual value of the i-th load of the b-th sample in the batch at time step t, is the predicted value of the i-th load of the b-th batch of samples at time step t; index i represents the i-th load forecasting task, index t is the time step, and the value range is 1 to 24; index b represents the sample index within the batch, and the value range is 1 to B;

[0019] Each load forecasting task is equally important, so the prediction accuracy loss function for each batch is as follows:

[0020]

[0021] Economic dispatch cost loss: After obtaining the load forecast results of the previous day from each dual encoder-single decoder, first add a dummy dimension to each [batch size × 24] matrix in the third dimension to form a [batch size × 24 × 1] matrix; then, merge in the third dimension in the order of the input sequence of learning-optimization to create a new three-dimensional matrix of size [batch size × 24 × n]; then input it into the pre-trained learning-optimization model to obtain a matrix of size [batch size × 1], which represents the economic dispatch cost calculated using the predicted load;

[0022] For actual values, follow the same steps to obtain the actual day-ahead economic dispatch cost:

[0023]

[0024] in, and i are the predicted value and actual value matrices of the i-th load forecasting task, both of size Add a pseudo dimension to convert them and i ', both sizes are The concatenated matrix of predicted and actual values, and y, of size Predicted economic dispatch cost obtained by learning-optimization model and the actual economic dispatch cost C ED The size of

[0025] The loss function of each batch of economic dispatch is as follows:

[0026]

[0027] If the predicted day-ahead economic dispatch cost value is lower than the day-ahead actual value, a penalty factor of 5 times will be imposed;

[0028] The prediction accuracy loss and economic dispatch loss are equally important, and the final loss function designed is:

[0029] Loss = loss of accuracy + loss of economic scheduling.

[0030] The beneficial effects of the present invention are: (1) The present invention adopts a customized encoder-decoder structure and combines deep learning models such as LSTM and CNN to effectively extract time series and spatial features, thereby significantly improving the accuracy of load forecasting.

[0031] (2) In large-scale power systems, the computational efficiency of load forecasting models is crucial. Traditional deep learning models usually have problems such as long training time and high consumption of computing resources when processing large-scale data.

[0032] (3) By designing a unified loss function, load forecasting is combined with dispatching objectives, avoiding the problem of increased dispatching costs caused by the inconsistency between forecast error and dispatching objectives in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the present invention.

[0034] Figure 2It is the comparison result of the predicted scheduling cost of the present invention applied to the test system (IEEE-9 node, IEEE-39 node) and other models. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0036] like Figure 1 As shown, an end-to-end multi-task load forecasting method for day-ahead economic dispatch includes the following steps:

[0037] S1. According to the standard day-ahead economic dispatch algorithm, the economic dispatch problem is constructed and solved, and then the input-output relationship of the neural network is trained to obtain the learning-optimization model, which is saved as a pre-trained model;

[0038] In the power system, economic dispatch (ED) is a key task. Its core goal is to ensure that the system can achieve efficient operation at the lowest cost while meeting the power demand by optimizing the allocation and utilization of power resources. The ED solution process is shown below.

[0039]

[0040] In the formula represents the active power generation of bus i at time t; represents the active power demand of bus i at time t; represents the reactive power of bus i at time t; represents the reactive power demand of bus i at time t; v i (t) is the voltage amplitude of bus i at time t; θ i (t) is the voltage angle of bus i at time t. In addition, represents the active power flow of the transmission line between bus i and bus j at time t, It represents the reactive power flow of the transmission line between bus i and bus j at time t. and Defines the minimum and maximum values ​​of the bus i voltage. is the maximum allowable voltage angle difference between bus i and bus j. Parameters and Respectively represent the minimum and maximum active power generation of bus i, and the parameters and They represent the minimum and maximum reactive power generation of bus i respectively. is the maximum apparent power flow of the transmission line between bus i and bus j. ij and b ijare the conductance and susceptance of the transmission line between bus i and bus j. and They represent the maximum rising and falling ramp rates of the active generation of bus i, respectively. Sets N, E, and T represent the set of all buses in the network, the set of all transmission lines in the network, and the set of all time periods considered in the optimization.

[0041] In the power system, we need to That is, the active power demand (active load) of bus i at time t, and That is, the reactive power (reactive load) of bus i at time t, formulate a dispatch plan, we take the IEEE 9-node (9-bus) system as an example, each BUS (bus) in the power system has its specific function and role. They are interconnected through generators, loads and transmission lines to form a complete small power system. The following is a detailed explanation of each BUS in the IEEE 9-bus system:

[0042] Bus 1, 2, 3 - Generation bus

[0043] Type: Generator bus (also called "reference bus" or "balance bus")

[0044] Function: Busbars 1, 2, and 3 are usually set as the reference point of the system, and are all connected to a generator (synchronous generator). It is used as the reference voltage of the system to maintain the frequency and voltage stability of the power system. It provides active power and reactive power of the system, regulates frequency, and serves as the core of power balance.

[0045] Bus 4, 5, 6 - load bus

[0046] Type: Load bus

[0047] Function: Buses 4, 5, and 6 are connected to both active and reactive loads. This bus is a major load point that receives power from buses 1, 2, and 3 and distributes it to the connected loads. These loads can represent industrial users or commercial areas.

[0048] Bus 7, 8, 9 - Intermediate transmission bus

[0049] Type: Intermediate bus

[0050] Function: Busbars 7, 8, and 9 are all intermediate transmission buses, used to connect the power generation busbar and other load buses. They transmit electricity from the power generation end to the load end through the transmission line, ensuring the transmission and distribution of electricity in the system.

[0051] Each busbar is connected to each other through transmission lines and transformers to form a complete small power network. By simulating the power flow and fault conditions between these busbars, the stability, reliability and various control strategies of the power system can be studied.

[0052] According to the above optimization problem, the MATPOWER optimizer is used to solve the above optimization problem. For an input matrix Load of size [N*24] (N represents the number of loads in a power system (including active loads and reactive loads on all load buses), and 24 represents 24 time points in a day), the output value Cost is obtained through the AC-OPF-based optimizer, with a size of [1*1], which represents the cost required for power system scheduling.

[0053] Through the above solver, we obtain a mapping from the load matrix of the input power system to the day-ahead dispatch cost, which corresponds to the input and output of the neural network. The purpose of training the neural network is to accurately predict the dispatch cost directly through the input load matrix through the trained neural network without going through a complicated solution process.

[0054] After obtaining the optimizer input-output relationship, the neural network training input-output relationship is established and saved as a pre-trained model. At the same time, in the learning-optimization model, the present invention uses the transposition learning method to establish input-output mapping from multiple angles, effectively improving the fitting ability.

[0055] S2. Design a deep learning structure of dual encoders and single decoders based on the sequence-to-sequence algorithm. Use two encoders to learn the dependencies and features in the time series data respectively, and fuse the information into the decoder for unified decoding.

[0056] In time series forecasting, although the accuracy of one-step forecasting is usually very high, the accumulation of forecast errors and uncertainties as the forecast time step increases poses a challenge to improving accuracy. Traditional multi-step forecasting usually adopts step-by-step iteration or direct multi-output methods, both of which are plagued by major problems. Since each step depends on the prediction of the previous time step, one-step iteration leads to a large amount of error accumulation. Although each step is independent in the direct multi-output method, it ignores the interdependence between the outputs of each time step. The sequence-to-sequence method (Seq2Seq) uses an encoder to encode the time series data and generates the output sequence step by step through a decoder. When predicting each step, the model considers the information of all previous steps to solve these problems. Based on the Seq2Seq theory, this paper proposes a customized multi-step forecast encoder-decoder structure to further improve the accuracy of the previous day (24 steps) load forecast. Two encoders are used to describe the long-distance correlation and the correlation of features in the time series data, respectively, and then decoded by the same decoder. At the same time, more supplementary information is added to the decoder through data decomposition technology, which further improves the accuracy of multi-step load forecasting. In this invention, the encoder part undertakes the task of extracting high-level features of the input data. To this end, we introduce two encoders to encode time series features and spatial features respectively, which are used to capture the temporal dependency and spatial features in the data respectively. The specific design and significance of the two encoders and one decoder are as follows:

[0057] Encoder 1 for temporal relationship learning: The first encoder, which focuses on temporal relationship learning, uses a long short-term memory (LSTM) layer to process the input time series data and capture temporal dependencies and long-term dependencies. It uses a BiLSTM structure to capture information from both the forward and reverse directions, enriching the feature set. An attention mechanism is integrated to focus on important time steps, enhancing the model's ability to focus on key moments. The cell state and hidden state of the LSTM unit are stored and used as input to the decoder.

[0058] Encoder 2 for feature extraction learning: The second encoder is dedicated to feature extraction learning and uses a CNN (convolutional neural network) structure to extract local features from the time series. Each filter in the CNN is responsible for capturing features of different scales and patterns. The extracted features are then input into the Transformer encoder, which further enhances the feature representation and captures more complex feature patterns.

[0059] Decoder structure: The decoder first decomposes the original data using ICEEMDAN (Improved Complete EnsembleEmpirical Mode decomposition with Adaptive Noise) to extract additional information to assist decoding, which is connected to the feature output of the first encoder to enrich the decoder input with time dependencies and extracted features. The decoder uses an LSTM layer to convert the concatenated features into an output sequence. The decoded features are then further processed using a Transformer decoder to capture complex temporal and feature relationships from both encoders. Finally, a fully connected layer transforms the processed features into the final prediction results. For the proposed encoder-decoder structure, the input is historical payload data The output is the load forecast result for 24 steps (one day in the future)

[0060] Assuming that the entire power system has N loads that need to be predicted (including active and reactive power), I need to train N of the above encoder-decoder structures in parallel. Each encoder-decoder corresponds to a load that needs to be predicted, and optimize them uniformly through a LOSS. Finally, I can get the prediction results of all loads in the power system at the same time.

[0061] S3. Based on the learning-optimization model and the encoder-decoder structure, a parallel load forecasting framework is designed. Multi-task learning is performed through the parallel encoder-decoder structure. At the same time, the accuracy of economic dispatch is considered on the basis of prediction accuracy through a customized loss function and a pre-trained learning-optimization model, making the load forecasting result more suitable for the day-ahead economic dispatch of the power system.

[0062] In order to achieve the unity of prediction and dispatching objectives, the present invention designs a loss function that integrates load prediction error and dispatching cost. The loss function consists of two parts: prediction accuracy loss and economic dispatching cost loss, as follows:

[0063] Prediction accuracy loss: At the beginning of training, all different load data are input into the corresponding codec structure for load prediction, and the prediction results for multiple time steps (the day before (24 steps)) are output. Here, each codec structure is parallel and independent, and the number of codec structures is related to the scale of the power system, that is, the number of codec structures is equal to the number of load prediction tasks in the power system (for example, in a 9-bus system, there are 6 load prediction tasks, so there are 6 parallel codec structures in the present invention). For each batch, the output matrix size of each encoder-decoder structure is [batch size × 24]. We assume that there are n different loads in the power system, and the prediction accuracy loss of each load prediction task in each batch is:

[0064]

[0065] Here, B represents the batch size, that is, the number of samples in each batch. The number 24 represents the number of time steps, corresponding to 24 hours in a day. is the actual value of the i-th load of the b-th sample in the batch at time step t, is the predicted value of the i-th load of the b-th sample at time step t. Index i represents the load forecasting task, and there are n load forecasting tasks in total. Index t is the time step, ranging from 1 to 24. Index b represents the sample index within the batch, ranging from 1 to b. In this invention, we consider each load forecasting task to be equally important. Therefore, the prediction accuracy loss function for each batch is as follows:

[0066]

[0067] Economic dispatch cost loss: After obtaining the load forecast results for the previous day (24 steps) from each encoder / decoder, we first add a dummy dimension to each [batch size × 24] matrix in the third dimension to form a [batch size × 24 × 1] matrix. Then, we merge them in the third dimension in the order of the input sequence of the learning-optimization to create a new three-dimensional matrix of size [batch size × 24 × n]. This is then input into the pre-trained learning-optimization model to obtain a matrix of size [batch size × 1], representing the economic dispatch cost calculated using the forecasted load. For the actual value, we follow the same steps to obtain the actual day-ahead economic dispatch cost.

[0068]

[0069] in, and i are the predicted value and actual value matrices of the i-th load forecasting task, both of size Add a pseudo dimension to convert them and i ', both sizes are The concatenated matrix of predicted and actual values, and y, of size Predicted economic dispatch cost obtained by learning-optimization model and the actual economic dispatch cost C ED The size of

[0070] After obtaining the predicted day-ahead dispatch cost and the actual day-ahead dispatch cost, we use the improved mean square error loss function to measure the difference between them. Since the penalty cost caused by undergeneration is higher than that caused by overgeneration, we need to avoid the situation where the predicted economic dispatch cost is lower than the actual economic dispatch cost. This problem is solved by the penalty loss technique. The economic dispatch loss function for each batch is as follows:

[0071]

[0072] If the predicted day-ahead economic dispatch cost value is lower than the day-ahead actual value, we will impose a penalty factor of 5 times.

[0073] We believe that prediction accuracy loss and economic dispatch loss are equally important, so the final loss function we designed is:

[0074] Loss = accuracy loss + economic scheduling loss

[0075] After the above model training is completed, the trained dual encoder-decoder is directly used to realize load forecasting, and based on the load forecasting results, the standard day-ahead economic dispatch algorithm can be used to solve the optimal day-ahead economic dispatch plan through the optimizer.

[0076] The present invention adopts a customized encoder-decoder structure and combines deep learning models such as LSTM and CNN to effectively extract time series and spatial features, so that the accuracy of load forecasting is significantly improved. Traditional load forecasting methods can usually only process data of a single dimension, such as only considering time series information or spatial information, and it is difficult to fully capture the complex relationships in power load data. The multi-task learning framework in the present invention can accurately capture load change trends at different time points and spatial locations by processing multiple prediction tasks in parallel.

[0077] Experimental results show that the prediction error of the method of the present invention is significantly lower than that of the traditional method in the application of IEEE 9-BUS and 39-BUS systems. It shows better performance in indicators such as MSE (mean square error) and MAE (mean absolute error). Especially in periods with large load fluctuations, such as peak and trough periods, the method of the present invention can more accurately predict load changes and ensure the effectiveness of the scheduling plan.

[0078] In large-scale power systems, the computational efficiency of load forecasting models is crucial. Traditional deep learning models usually have problems with long training time and high consumption of computing resources when processing large-scale data. The present invention significantly improves the training speed of the model by introducing accelerated training technology based on learning-optimization. Learning-optimization technology greatly reduces training and inference time through efficient feature fusion and normalization processing. This improvement in efficiency enables the present invention to process prediction tasks of large-scale power systems in a shorter time, making it more suitable for load forecasting and scheduling scenarios.

[0079] The present invention combines load forecasting with dispatching objectives by designing a unified loss function, thus avoiding the problem of increased dispatching costs caused by the inconsistency between forecasting errors and dispatching objectives in traditional methods. Specifically, by integrating dispatching costs into the loss function, the present invention can not only minimize the forecasting error, but also directly optimize the dispatching strategy, ensuring that the generated dispatching plan is more economical and efficient.

[0080] In the experiment, the method of the present invention showed significant reduction in dispatching costs in multiple test scenarios. In the tested IEEE 39-BUS system, the traditional dispatching method usually generates additional dispatching costs in actual operation due to prediction deviation, such as starting and stopping more generators or making more power transmission adjustments. The present invention effectively reduces these additional operations through more accurate load forecasting and joint optimization with dispatching tasks, and the dispatching cost is reduced by about 15%-25% on average.

[0081] The operating environment of the power system is highly uncertain, especially when facing emergencies (such as extreme weather, equipment failure, etc.), load fluctuations and scheduling requirements are more complex and changeable. The method of the present invention enhances the robustness and adaptability of the model through the design of a multi-task learning framework.

[0082] By processing multiple load forecasting tasks in parallel, the present invention can maintain high forecasting accuracy and scheduling efficiency under different load fluctuation scenarios. For example, in the experiment, when the load of the power system fluctuates violently, the method of the present invention can still effectively cope with it, while the prediction error of the traditional method increases significantly. In addition, the present invention uses a customized encoder-decoder structure to enable the model to adapt to power systems of different sizes, whether it is a small-scale regional power grid or a large-scale national power system, showing good adaptability.

[0083] Traditional dispatching methods often lead to instability in dispatching strategies due to the disconnection between prediction errors and dispatching targets, especially when load forecasts are inaccurate or dispatching demand changes greatly, which easily leads to frequent dispatching adjustments and waste of resources. The present invention combines dispatching targets in the loss function to generate a more stable and reliable dispatching strategy.

[0084] The method of the present invention has not only proved its effectiveness in theory, but also demonstrated wide adaptability in practical applications. Through experimental verification, the method of the present invention has shown excellent performance in different types of power systems, including regional power grids, urban power grids and large-scale national power grid systems. In addition, the method can be seamlessly integrated with the existing power dispatching system, greatly improving the intelligence and automation of the system. By combining the powerful feature extraction capability of deep learning with the unified framework of dispatching optimization objectives, the present invention provides an efficient, intelligent and economical solution for the development of future smart grids. The flexibility and scalability of the method enable it to maintain good adaptability when facing more complex power dispatching needs in the future.

[0085] In an embodiment of the present application, during the training phase of the model, after data preprocessing, the standardized data will be input into the feature extraction module. The encoder-decoder structure extracts complex features through a multi-layer neural network, and achieves unified optimization of load forecasting and scheduling costs by optimizing the loss function. The model is trained in batches to reduce the occupation of computing resources, and the optimal model parameters are determined by cross-validation.

[0086] In practical applications, the model can receive real-time load data in the power system and generate optimal dispatch strategies based on the prediction results. These strategies can effectively reduce the operating costs of the power system and improve the efficiency and reliability of the overall system.

[0087] The present invention uses a system containing IEEE 9 nodes and IEEE 39 nodes, and uses real net load data from Australia to perform simulations, and obtains the following results as shown in Table 1 and Figure 2 The forecast dispatch cost is compared with other models. The compared models are using LSTM for separate prediction, using encoder-decoder structure for separate prediction, using LSTM for end-to-end prediction, and comparing the forecast economic dispatch cost calculated based on the load forecast results with the actual economic dispatch cost. The evaluation indicators used are mean absolute error (MAPE), mean square error (MSE), mean absolute percentage error (MAE), and determination coefficient (R2).

[0088] Table 1 Comparison results of the predicted scheduling cost in the test system with other models.

[0089]

[0090] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. An end-to-end multi-task load forecasting method for day-ahead economic dispatch, characterized by: The following steps are involved: S1. According to the standard day-ahead economic dispatch algorithm, the economic dispatch problem of the power system is constructed and solved, and then the input-output relationship of the neural network training is established to obtain the learning-optimization model and save it as a pre-training model; S2. Design a deep learning structure of dual encoders and single decoders based on the sequence-to-sequence algorithm. Use two encoders to learn the dependencies and features in the time series data respectively, and fuse the information into the decoder for unified decoding. S3. Based on the learning-optimization model and the encoder-decoder structure, a parallel load forecasting framework is designed. Multi-task learning is performed through the parallel encoder-decoder structure. At the same time, the accuracy of economic dispatch is considered on the basis of prediction accuracy through the designed loss function combined with the learning-optimization model.

2. The end-to-end multi-task load forecasting method for day-ahead economic dispatch according to claim 1, characterized in that: In step S1, the power system includes a generator, a load, a transmission line, and a plurality of buses for connecting the generator, the load, and the transmission line, wherein the bus includes a power generation bus, a load bus, and an intermediate transmission bus; The standard day-ahead economic dispatch algorithm implements the allocation and utilization of power resources based on the active and reactive loads of each load bus at time t, thereby ensuring that the system can achieve efficient operation at the lowest cost while meeting power demand: For an input matrix Load of size [N*24], N represents the number of loads in a power system, including active loads and reactive loads on all load buses, and 24 represents 24 time points in a day. The output value Cost is obtained by the AC-OPF-based optimizer, with a size of [1*1], which represents the cost required for power system scheduling; After the neural network is established, the input matrix Load is used as the neural network input, and the output value Cost corresponding to the input matrix is ​​used as the expected input of the neural network. That is, a learning-optimization model is established and saved as a pre-trained model.

3. The end-to-end multi-task load forecasting method for day-ahead economic dispatch according to claim 1, characterized in that: The dual encoder-single decoder deep learning structure includes: The first encoder focuses on temporal relationship learning and uses a long short-term memory layer to process input time series data to capture temporal dependencies and long-term dependencies. The encoder uses a BiLSTM structure to capture information from both the forward and reverse directions, enriches the feature set, and integrates an attention mechanism to focus on important time steps, enhancing the model's ability to focus on key moments. The second encoder is dedicated to feature extraction learning and uses a CNN network structure to extract local features from the time series. Each filter in the CNN is responsible for capturing features of different scales and patterns. The extracted features are then input into the Transformer encoder, which further enhances the feature representation and captures more complex feature patterns. The decoder first uses ICEEMDAN to decompose the original data and extract additional information to assist decoding. This information is connected to the feature output of the first encoder, enriching the decoder input with temporal dependencies and extracted features. The decoder uses an LSTM layer to convert the concatenated features into an output sequence, and then uses a Transformer decoder to further process the decoded features, capturing complex temporal and feature relationships from the two encoders. Finally, a fully connected layer converts the processed features into the final prediction results.

4. The end-to-end multi-task load forecasting method for day-ahead economic dispatch according to claim 1, characterized in that: Assuming that the power system has N load buses that need to predict loads, for each load bus, a dual encoder-single decoder structure described in step S2 needs to be constructed; When training the dual encoder-single decoder structure for each load bus, firstly, based on the historical load data, historical load data of a specified time length is selected as the input of the dual encoder-single decoder structure, and the dual encoder-single decoder structure outputs the prediction result.

5. The end-to-end multi-task load forecasting method for day-ahead economic dispatch according to claim 1, characterized in that: In step S3, at the beginning of training, historical load data of a specified time length is selected for each load bus and input into the corresponding dual encoder-single decoder structure for load prediction, and each dual encoder-single decoder structure outputs prediction results of multiple time steps; then the prediction results of multiple time steps are input into the learning-optimization model to obtain the prediction results of the lowest scheduling cost; and on this basis, a loss function is designed, and based on the loss function, the training update of the learning-optimization model and each dual encoder-single decoder structure is realized; In this process, a loss function that unifies load forecasting and dispatching objectives is established. This loss function not only focuses on the accuracy of the forecast, but also directly takes the dispatching cost as part of the optimization objective, including: For each batch prediction, the output matrix size of each dual encoder-decoder structure is [batch size × 24]. Assuming that there are n different load forecasting tasks in the power system, n = N, each load forecast corresponds to a load bus, and the predicted load is also called the load data of the load bus. The prediction accuracy loss of each batch prediction of each load forecasting task is: Among them, B represents the batch size, that is, the number of samples in each batch; the number 24 represents the number of time steps, corresponding to 24 hours in a day; is the actual value of the i-th load of the b-th sample in the batch at time step t, is the predicted value of the i-th load of the b-th batch of samples at time step t; index i represents the i-th load forecasting task, index t is the time step, and the value range is 1 to 24; index b represents the sample index within the batch, and the value range is 1 to B; Each load forecasting task is equally important, so the prediction accuracy loss function for each batch is as follows: Economic dispatch cost loss: After obtaining the load forecast results of the previous day from each dual encoder-single decoder, first add a dummy dimension to each [batch size × 24] matrix in the third dimension to form a [batch size × 24 × 1] matrix; then, merge in the third dimension in the order of the input sequence of learning-optimization to create a new three-dimensional matrix of size [batch size × 24 × n]; then input it into the pre-trained learning-optimization model to obtain a matrix of size [batch size × 1], which represents the economic dispatch cost calculated using the predicted load; For actual values, follow the same steps to obtain the actual day-ahead economic dispatch cost: in, and i are the predicted value and actual value matrices of the i-th load forecasting task, both of size Add a pseudo dimension to convert them and i ', both sizes are The concatenated matrix of predicted and actual values, and y, of size Predicted economic dispatch cost obtained by learning-optimization model and the actual economic dispatch cost C ED The size of The loss function of each batch of economic dispatch is as follows: The prediction accuracy loss and economic dispatch loss are equally important, and the final loss function designed is: Loss = loss of accuracy + loss of economic scheduling.

Images