Power short-term load prediction system based on multi-source information adaptive fusion and working method thereof

Abstract

The invention provides a power short-term load prediction system based on multi-source information adaptive fusion and a working method thereof, and the system comprises a multi-view data construction module which is used for organizing heterogeneous input data into a plurality of independent semantic views; the spatial dependence coding module is used for learning spatial dependence among nodes through a graph attention network and generating spatial embedding representation; the time dynamic modeling module is used for processing the time sequence view through a recurrent neural network and generating time feature embedded representation; the auxiliary information modeling module is used for processing other auxiliary views through independent feedforward neural network processing and generating auxiliary feature embedded representation; the cross-view attention fusion module is used for carrying out adaptive weighted fusion on all the embedded representations so as to inhibit representation degradation caused by low-quality views; and the load prediction module is used for generating a load prediction result based on the fusion representation and carrying out end-to-end training by taking a mean square error as a target.

Description

TECHNICAL FIELD

[0001] The application provides a power short-term load forecasting system based on adaptive fusion of multi-source information and a working method thereof, and relates to the technical fields of machine learning, deep learning and smart grid. BACKGROUND

[0002] Short-term power load forecasting is crucial for the reliable operation of power grids and the integration of renewable energy. With the development of smart grids, the data available for forecasting has increased, including historical load, meteorological data, socio-economic indicators, and power grid topology.

[0003] Existing spatio-temporal models usually adopt a simple method when dealing with these heterogeneous data, which is to concatenate all different features (such as time series, graph attributes, auxiliary features) in the input layer to form a single input representation. This simple concatenation strategy has significant drawbacks: it ignores the inherent semantic differences and complementarities between different data sources. More seriously, it risks "representation degradation": it forces all information to be compressed into the same representation space, which can cause low-quality or noisy views to weaken the representational capacity of high-quality views, resulting in a decline in overall representation quality and ultimately harming forecasting performance.

[0004] Although techniques such as graph convolutional networks, graph attention networks, and long short-term memory networks have been used for spatio-temporal modeling, they are still mostly limited to processing this concatenated single input and have not fundamentally addressed the challenge of heterogeneous information fusion. Therefore, there is an urgent need for a new framework that can explicitly separate different information sources and adaptively balance their contributions to achieve more effective fusion. SUMMARY

[0005] In view of this, in order to make up for the gaps and deficiencies in the prior art, the application provides a power short-term load forecasting system based on adaptive fusion of multi-source information and a working method thereof, which can accurately and effectively perform short-term power load forecasting.

[0006] The application provides a power short-term load forecasting system based on adaptive fusion of multi-source information and a working method thereof, which includes the following contents:

[0007] The application provides a power short-term load forecasting system based on adaptive fusion of multi-source information, characterized by the following contents: the power short-term load forecasting system based on adaptive fusion of multi-source information includes a multi-view data construction module, a spatial dependency encoding module, a temporal dynamic modeling module, an auxiliary information modeling module, a cross-view attention fusion module, and a load forecasting module.

[0008] The multi-view data construction module is used for organizing heterogeneous input data including historical load, weather, social economic indicators and power grid topology into multiple independent semantic views.

[0009] The spatial dependence coding module is used for receiving a graph structure view of the power grid, learning spatial dependence between nodes through a graph attention network, and generating a spatial embedding representation of the nodes.

[0010] The time dynamic modeling module is used for processing a time series view through a recurrent neural network, and generating a time feature embedding representation.

[0011] The auxiliary information modeling module is used for processing other auxiliary views, and converting the auxiliary views into auxiliary feature embedding representations through independent feedforward neural networks.

[0012] The cross-view attention fusion module is used for receiving the spatial, time and auxiliary embedding representations, adaptively assigning weights to the representations of different views through a learnable attention mechanism, and weightedly fusing the representations into a unified feature representation.

[0013] The load prediction module is used for outputting a power load prediction value at a future time point based on the fused feature representation, and updating parameters of the entire model through end-to-end training by minimizing a mean square error loss function between the prediction value and an actual value, and outputting a power load prediction result corresponding to the to-be-measured node based on an optimal model if the loss converges to a stable value.

[0014] The application further provides a working method of the power short-term load prediction system based on adaptive fusion of multi-source information.

[0015] Heterogeneous input data including historical load, weather, social economic indicators and power grid topology are taken as input, and are firstly organized into multiple independent semantic views through the multi-view data construction module; then the corresponding views are encoded through the spatial dependence coding module, the time dynamic modeling module and the auxiliary information modeling module respectively to extract respective feature embedding representations; all the embedding representations are adaptively weightedly fused through the cross-view attention fusion module to generate a unified feature representation; finally, the optimal model is obtained through iterative training of the load prediction module, and the final power load prediction result is obtained by using the optimal model.

[0016] Further, the working method of the multi-view data construction module further comprises: receiving heterogeneous power system data, the data comprising but not limited to hourly load, renewable energy generation, meteorological variables, power grid topology, socio-economic indicators and calendar features; reorganizing the heterogeneous data into a plurality of independent semantic views, in a specific embodiment, the views are organized into: a load and renewable energy view comprising hourly load and renewable energy generation data; a meteorological view comprising meteorological variable data; a graph view comprising static power grid topology and node and edge attributes; a socio-economic view comprising GDP, installed capacity and population data; and a calendar view comprising features extracted from timestamps.

[0017] Further, the working method of the spatial dependency encoding module further comprises:

[0018] First, receiving the graph view, wherein node i has an initial feature representation g i , and the edge between node i and j has an attribute vector e ij ; the edge attribute vector e ij is first projected into a latent space by a learnable projection matrix U to obtain e′ ij = Ue ij ;

[0019] Then, the attention coefficient of node i and its neighbor on the kth attention head is calculated , and the calculation formula is:

[0020]

[0021] Where W (k) and a (k) are the learnable weight matrix and attention vector of the kth head, respectively, σ(·) is a nonlinear activation function, and || represents the concatenation operation; finally, the spatial embedding representation of the node is obtained by aggregating through H attention heads, and the calculation formula is:

[0022]

[0023] Where is the spatial embedding representation of the node.

[0024] Further, the working method of the time dynamic modeling module further comprises:

[0025] First, for any time series view of node i , concatenate the feature at the historical time step τ with the spatial embedding representation generated by the spatial dependency encoding module to obtain a combined input The calculation formula is:

[0026]

[0027] Then, the combination input sequence is input into a recurrent neural network for processing, and the recursive function F LSTM is represented as:

[0028]

[0029] wherein and are hidden states and cell states, respectively, and Θ LSTM is a learnable parameter; finally, the hidden state of the recurrent neural network at the last time step t is taken as the temporal feature embedding representation of the temporal view, and the calculation formula is:

[0030]

[0031] wherein is the temporal feature embedding representation of the temporal view.

[0032] Further, the working method of the auxiliary information modeling module further includes:

[0033] First, the auxiliary view

[0034] is received; then, the auxiliary feature of node i at the last time step t (v) is transformed through a feedforward neural network MLP specific to the view; finally, the transformed result is output as the feature embedding representation of the auxiliary view , and the calculation formula is:

[0035]

[0036] wherein MLP (v) is a feedforward neural network with independent parameters.

[0037] Further, the working method of the cross-view attention fusion module further includes:

[0038] First, since the embeddings from different encoding modules may be different in dimension, they are projected to a unified common latent space through a view-specific learnable projection matrix P (v) , to obtain , and the calculation formula is:

[0039]

[0040] Then, a learnable weight parameter w is assigned to each of the V views v v And it is normalized by the Softmax function to obtain the adaptive view importance weight The calculation formula is:

[0041]

[0042] Finally, the projection embedding of all views According to the corresponding weight Weighted sum, the final unified feature representation is obtained, and the calculation formula is:

[0043]

[0044] Wherein The final unified feature representation.

[0045] Further, the working method of the load prediction module further comprises:

[0046] First, the unified feature representation output by the cross-view attention fusion module is received Then, the fusion feature representation Is sent into a prediction layer Predictor(·) to generate the power load prediction value of the next moment t+1 The calculation formula is:

[0047]

[0048] Finally, by minimizing the mean square error loss function Loss between the prediction value And the real load value y i,t+1 The entire model is trained end-to-end, and the loss function formula is:

[0049]

[0050] Wherein N is the number of nodes in the batch, if the loss L converges to a stable value, the power load prediction result corresponding to the to-be-measured node is output based on the optimal model.

[0051] The present application has the following advantages:

[0052] In view of the problem that the existing model simply splices heterogeneous features, resulting in loss of semantic information and degradation of representation, the present application designs a multi-view data construction module, which explicitly organizes data into multiple independent semantic views, and processes them through dedicated coding modules, i.e. spatial dependence coding module, time dynamic modeling module and auxiliary information modeling module, to realize flexible integration of multi-source information.

[0053] In view of the challenge that low-quality or noisy views may weaken the overall representation quality in multi-view learning, the application designs a cross-view attention fusion module. The module adaptively assigns weights to the representations of different views through a learnable attention mechanism, effectively suppresses representation degradation, and more fully utilizes the complementary information between different data sources.

[0054] The application can simultaneously and explicitly capture the spatial dependence in the power network topology and the time series dynamics of the load data by combining the graph attention network to process the spatial views and combining the recurrent neural network to process the time views, and effectively models the complex spatio-temporal dependence relationship.

[0055] The scheme can efficiently integrate spatio-temporal data, meteorological, social and economic and other multi-source heterogeneous information, and accurately and effectively perform short-term power load forecasting. Experimental results show that the scheme proposed in the application has a significant improvement in prediction accuracy compared with the prior art, verifying the effectiveness of the application. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 The scheme of the application is implemented.

[0057] Figure 2 The neural network architecture of the system of the application is shown. DETAILED DESCRIPTION

[0058] The technical scheme of the application will be specifically described below with reference to the drawings.

[0059] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0060] It should be noted that the terms used herein are only for the purpose of describing the specific embodiments and are not intended to limit the exemplary embodiments according to the application; as used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.

[0061] As shown in Figures 1 to 2 The application proposes a power short-term load forecasting system based on adaptive fusion of multi-source information and a working method thereof, including the following contents:

[0062] The application proposes a power short-term load forecasting system based on adaptive fusion of multi-source information, characterized by including the following contents:

[0063] The application provides a power short-term load prediction system based on adaptive fusion of multi-source information, and has the characteristics that the system comprises the following contents: the power short-term load prediction system based on adaptive fusion of multi-source information comprises a multi-view data construction module, a spatial dependence coding module, a time dynamic modeling module, an auxiliary information modeling module, a cross-view attention fusion module and a load prediction module;

[0064] The multi-view data construction module is used for organizing heterogeneous input data comprising historical load, meteorology, social economic indicators and power grid topology into multiple independent semantic views.

[0065] The spatial dependence coding module is used for receiving a graph structure view of a power grid, learning spatial dependence relationships between nodes through a graph attention network, and generating spatial embedding representations of the nodes.

[0066] The time dynamic modeling module is used for processing a time sequence view through a recurrent neural network, and generating time feature embedding representations.

[0067] The auxiliary information modeling module is used for processing other auxiliary views, and converting the auxiliary views into auxiliary feature embedding representations through independent feedforward neural networks.

[0068] The cross-view attention fusion module is used for receiving the spatial, time and auxiliary embedding representations, adaptively assigning weights to the representations of different views through a learnable attention mechanism, and weightedly fusing the representations into a unified feature representation.

[0069] The load prediction module is used for outputting power load prediction values at future time points through a prediction layer based on the fused feature representation, and training and updating parameters of the entire model in an end-to-end manner by minimizing a mean square error loss function between the prediction values and actual values, so that if the loss converges to a stable value, the power load prediction result corresponding to the to-be-measured node is output based on the optimal model.

[0070] The application provides a working method of the power short-term load prediction system based on adaptive fusion of multi-source information, and a corresponding typical working process is as shown in Figure 1 and Figure 2 The working process can be regarded as the following steps:

[0071] Step S1: heterogeneous input data comprising historical load, meteorology, social economic indicators and power grid topology are input into the multi-view data construction module, and the data are organized into multiple independent semantic views, such as a graph view, a load time sequence view, a meteorology view and the like.

[0072] Step S2: the graph view constructed in step S1 is input into the spatial dependence coding module to generate spatial embedding representations.

[0073] Step S3: input the temporal view constructed in step S1 into a temporal dynamics modeling module to generate a temporal feature embedding representation;

[0074] Step S4: input the auxiliary view constructed in step S1 into an auxiliary information modeling module to generate an auxiliary feature embedding representation.

[0075] Step S5: input the multiple feature embedding representations obtained in steps S2, S3 and S4 into a cross-view attention fusion module to generate a unified feature representation through adaptive weighted fusion.

[0076] Step S6: input the fused feature representation obtained in step S5 into a load prediction module to generate a prediction value through a prediction layer.

[0077] Step S7: calculate a mean square error loss function according to the prediction value and the true value obtained in step S6, and update the parameters of the model using back propagation and gradient descent algorithm according to the loss function. If the loss function does not converge to a stable value, return to step S2.

[0078] Step S8: if the loss function converges, an optimal model is obtained. According to the optimal model, the test set data is predicted, and the final power load prediction result is output.

[0079] In the present embodiment, step S1 specifically comprises the following steps:

[0080] Step S11: receive heterogeneous power system data.

[0081] Step S12: pre-process the data, for example, use forward and backward padding to interpolate missing values, and use min-max normalization to scale all numerical variables to the [0, 1] interval.

[0082] Step S13: reorganize the pre-processed heterogeneous data into multiple independent semantic views for subsequent modules. In one specific embodiment, the views are organized as: a load and renewable energy view including hourly load and renewable energy generation data; a meteorological view including meteorological variables; a graph view including static power grid topology and node and edge attributes; a socio-economic view including GDP, installed capacity and population data; and a calendar view including features extracted from timestamps such as hour, day of the week, month, holiday and season.

[0083] In the present embodiment, step S2 specifically comprises the following steps:

[0084] Step S21: receive the view constructed in step S1, wherein node i has an initial feature representation g i , and the edge between node i and j has an attribute vector e ij ;

[0085] Step S22: projecting the edge attribute vector e ij First, it is projected into latent space by a learnable projection matrix U, obtaining e ' ij = Ue ij ;

[0086] Step S23: calculating the node i and its neighbors The attention coefficient on the kth attention head is The calculation formula is:

[0087]

[0088] Where W (k) and a (k) are the learnable weight matrix and attention vector of the kth head respectively, σ(·) is a nonlinear activation function, and || represents the concatenation operation.

[0089] Step S24: aggregating through H attention heads to obtain the spatial embedding representation of the node The calculation formula is:

[0090]

[0091] In this embodiment, step S3 specifically includes the following steps:

[0092] Step S31: receiving the time series view v constructed in step S1 and obtaining its features at historical time steps τ = t-T+1, …, t.

[0093] Step S32: concatenating the time series features with the spatial embedding representation generated in step S2 to obtain a combined input The calculation formula is:

[0094]

[0095] Step S33: feeding the combined input sequence into a recurrent neural network for processing, and the recursive function F LSTM is represented as:

[0096]

[0097] Where and are the hidden state and cell state respectively, Θ LSTM is the learnable parameter.

[0098] Step S34: take the hidden state of the recurrent neural network at the last time step t as the time feature embedding representation of the temporal view and output it, whose computational formula is:

[0099]

[0100] In this embodiment, step S4 specifically comprises the following steps:

[0101] Step S41: receive the auxiliary view

[0102] Step S42: extract the auxiliary feature of node i at the last time step t

[0103] Step S43: send the feature into a view-specific feedforward neural network MLP (v) transform it, and output the transformed result as the feature embedding representation of the auxiliary view whose computational formula is:

[0104]

[0105] where MLP (v) is a feedforward neural network with independent parameters.

[0106] In this embodiment, step S5 specifically comprises the following steps:

[0107] Step S51: receive multiple feature embedding representations from step S2, step S3 and step S4 where v traverses all views V, including the graph view, the temporal view and the auxiliary view.

[0108] Step S52: since the embeddings from different encoding modules may be different in dimension, project them into a unified common latent space through a view-specific learnable projection matrix P (v) to obtain whose computational formula is:

[0109]

[0110] Step S53: assign a learnable weight parameter w to each of the V views v and normalize it through a Softmax function to obtain an adaptive view importance weight The weight is used to measure the contribution of different views to the final prediction, whose computational formula is:

[0111]

[0112] Step S54: Projection embedding of all views According to its corresponding weight Weighted sum to get the final unified feature representation And output, its calculation formula is:

[0113]

[0114] In this embodiment, step S6 specifically includes the following steps:

[0115] Step S61: Receive the unified feature representation from step S5

[0116] Step S62: Send the fusion feature representation Into a prediction layer Predictor(·) to generate the power load prediction value at the next time t+1 Its calculation formula is:

[0117]

[0118] In this embodiment, step S7 specifically includes the following steps:

[0119] Step S71: Receive the power load prediction value from step S6 And get the real load value y i,t+1 ;

[0120] Step S72: Calculate the mean square error loss function Loss between the prediction value And the real load value y i,t+1 End-to-end train the entire model, and the loss function formula is:

[0121]

[0122] Where N is the number of nodes in the batch.

[0123] Step S73: Determine whether the loss function converges to a stable value (for example, the validation set loss does not decrease again within a certain period), if not, return to step S2 for the next round of training iteration;

[0124] In this embodiment, step S8 specifically includes the following steps:

[0125] Step S81: According to the optimal model trained in step S7, perform load prediction on the power system nodes for testing, and output the final power load prediction result.

[0126] Particularly, the embodiment adopts a supervised learning training method. In view of the problem that simple splicing of heterogeneous features of an existing model leads to loss of semantic information and degradation of representation, a multi-view data construction module is designed to explicitly organize data into multiple independent semantic views. At the same time, a special coding module, i.e. a spatial dependence coding module, a time dynamic modeling module and an auxiliary information modeling module, is designed to process different types of data, realizing flexible integration of multi-source information. In view of the challenge that low-quality or noisy views may weaken the overall representation quality in multi-view learning, a cross-view attention fusion module is designed. The module adaptively assigns weights to the representations of different views through a learnable attention mechanism, can effectively suppress representation degradation, and more fully utilize the complementary information between different data sources. The short-term power load prediction model proposed in the application can efficiently integrate spatio-temporal data, weather, social and economic and other multi-source heterogeneous information, and accurately and effectively perform short-term power load prediction.

[0127] In an embodiment of the application, the model comparison based on prediction accuracy is shown in Table 1, including the following contents:

[0128] Model MAE(MW) RMSE (MW) MAPE (%) XGBoost 297.47 517.69 40.50 GCN-LSTM 89.11 184.12 5.72 EdgeGCNLSTM 84.63 148.09 7.24 LSTM 82.68 141.55 5.75 GAT-LSTM 64.64 119.06 4.59 MSIAF 56.48 99.54 4.37

[0129] Table 1 shows the model comparison based on prediction accuracy

[0130] Wherein as shown in Table 1, wherein based on the model comparison of prediction accuracy, the best result is the multi-source information adaptive fusion system (Multi-Source Information Adaptive Fusion System, referred to as MSIAF) proposed in the application. On the basis of the above design, the prediction performance of the present embodiment and other existing schemes is compared on a real power system data set of a certain region. In the experiment, the data set is divided into training set, validation set and test set in chronological order, wherein the training set is from January to December 2019; wherein the validation set is from January to June 2020; wherein the test set is from July to December 2020. The optimal model obtained after training is used to predict the power load of the test set, and the results of three performance indicators, i.e. MAE, RMSE and MAPE, are obtained. Among them, MAE (Mean Absolute Error) represents the average value of the absolute error between the predicted value and the true value, which can intuitively reflect the actual size of the prediction error; RMSE (Root Mean Square Error) represents the square root of the sum of squares of the deviation between the predicted value and the true value, which can measure the dispersion degree of the predicted value; MAPE (Mean Absolute Percentage Error) represents the average value of the absolute error percentage of the true value, which has the characteristic of dimensionless. It can be seen that the multi-source information adaptive fusion system (Multi-Source Information Adaptive Fusion System, referred to as MSIAF) proposed in the application has a significant performance improvement in all indicators compared with the existing baseline algorithms such as GAT-LSTM, GCN-LSTM and LSTM, which also verifies the effectiveness of the power short-term load prediction method in the application.

[0131] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROM, optical memory, etc.) having computer-usable program code embodied thereon.

[0132] The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0133] The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0134] The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can 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 such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0135] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the present application shall have the common meaning as understood by a person with ordinary skill in the art to which the present application belongs. The terms "first", "second" and similar terms used in the present application do not denote any order, quantity or importance, but are used to distinguish different components. The terms "include", "contain" and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connect" or "connected" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right" and the like are only used to represent relative positional relationships, and when the absolute positions of the described objects are changed, the relative positional relationships may

[0136] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application in other forms. Any person skilled in the art can make changes or modifications to the above disclosed technical contents to obtain equivalent embodiments. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application and according to the technical essence of the present application still falls within the protection scope of the present application.

[0137] The above are preferred embodiments of the present application. Any change made according to the technical solution of the present application and any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application and according to the technical essence of the present application still fall within the protection scope of the present application.

Claims

1. A short-term power load forecasting system based on multi-source information adaptive fusion, characterized in that, The following is included: The short-term power load forecasting system based on multi-source information adaptive fusion includes a multi-view data construction module, a spatial dependency coding module, a time dynamic modeling module, an auxiliary information modeling module, a cross-view attention fusion module, and a load forecasting module; The multi-view data construction module is used to organize heterogeneous input data, including historical load, meteorology, socioeconomic indicators and power grid topology, into multiple independent semantic views; The spatial dependency encoding module is used to receive a graph structure view of the power grid, learn the spatial dependencies between nodes through a graph attention network, and generate spatial embedding representations of the nodes. The time dynamic modeling module is used to process the time series view through a recurrent neural network and generate a time feature embedding representation; The auxiliary information modeling module is used to process other auxiliary views and convert them into auxiliary feature embedding representations through an independent feedforward neural network; The cross-view attention fusion module is used to receive the above spatial, temporal and auxiliary embedding representations, and adaptively assign weights to the representations of different views through a learnable attention mechanism, and then weightedly fuse them into a unified feature representation. The load prediction module is used to output the predicted power load value at future time points through a prediction layer based on the fused feature representation, and to train and update the parameters of the entire model end-to-end by minimizing the mean square error loss function between the predicted value and the true value. If the loss converges to a stable value, the power load prediction result corresponding to the node to be tested is output based on the optimal model.

2. A method for operating a short-term power load forecasting system based on multi-source information adaptive fusion, used in the short-term power load forecasting system based on multi-source information adaptive fusion as described in claim 1, characterized in that, Includes the following: Heterogeneous input data, including historical load, meteorological, socioeconomic indicators, and power grid topology, is used as input. First, the data is organized into multiple independent semantic views through a multi-view data construction module. Then, the corresponding views are encoded by spatial dependency encoding, temporal dynamic modeling, and auxiliary information modeling modules to extract their respective feature embedding representations. Next, a cross-view attention fusion module adaptively weights and fuses all the above embedding representations to generate a unified feature representation. Finally, the load forecasting module iteratively trains to obtain the optimal model, and the final power load forecast result is obtained using this optimal model.

3. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the multi-view data construction module further includes: receiving heterogeneous power system data, including but not limited to: hourly load, renewable energy generation, meteorological variables, grid topology, socioeconomic indicators, and calendar features; reorganizing the heterogeneous data into multiple independent semantic views. In one specific embodiment, the views are organized as: a load and renewable energy view, including hourly load and renewable energy generation data; a meteorological view, including meteorological variable data; a graph view, including static grid topology and node and edge attributes; a socioeconomic view, including GDP, installed capacity, and population data; and a calendar view, including features extracted from timestamps.

4. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the spatial dependency encoding module also includes: First, the graph view is received, where node i has an initial feature representation g. i The edge between nodes i and j has an attribute vector e. ij The edge attribute vector e ij First, it is projected into the latent space through a learnable projection matrix U, resulting in e′. ij =Ue ij ; Then, the multi-head attention mechanism is used to calculate the relationship between node i and its neighbors. Attention coefficient at the k-th attention head The calculation formula is as follows: Among them W (k) and a (k) These are the learnable weight matrix and attention vector of the k-th head, respectively, where σ(·) is the non-linear activation function and || represents the concatenation operation. Finally, the spatial embedding representation of the node is obtained by aggregating H attention heads, and its calculation formula is: in This is the spatial embedding representation of the node.

5. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the time dynamic modeling module also includes: First, for any time-series view of node i Its features at historical time step τ Spatial embedding representation generated by the spatial dependency encoding module By concatenating the components, a combined input is obtained. The calculation formula is as follows: Then, input the combined sequence. The data is fed into a recurrent neural network for processing, and its recursive function F... LSTM Represented as: in and These are the hidden state and the cell state, respectively. LSTM The parameters are learnable; finally, the hidden state of the recurrent neural network at the last time step t is taken. As the temporal feature embedding representation of the time-series view, its calculation formula is: in The temporal features of the time-series view are embedded and represented.

6. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the auxiliary information modeling module also includes: First, receive the auxiliary view. Then, through a view-specific feedforward neural network MLP (v) Auxiliary features of node i at the last time step t Perform the transformation; finally, output the transformed result as the feature embedding representation of the auxiliary view. The calculation formula is as follows: MLP (v) It is a feedforward neural network with independent parameters.

7. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the cross-view attention fusion module also includes: First, due to embedding from different encoding modules The dimensions may differ, and a specific learnable projection matrix P can be used to represent a view. (v) Projecting it onto a unified public potential space yields... The calculation formula is as follows: Then, a learnable weight parameter w is assigned to each of the V views v. v The view importance weights are then normalized using the Softmax function to obtain adaptive view importance weights. The calculation formula is as follows: Finally, through projection embedding of all views According to their corresponding weights Weighted summation yields the final unified feature representation, calculated using the following formula: in This represents the final unified feature representation.

8. The working method of a short-term power load forecasting system based on multi-source information adaptive fusion according to claim 2, characterized in that, The working method of the load forecasting module also includes: First, the unified feature representation output by the cross-view attention fusion module is received. Then, the fusion feature representation The data is fed into a prediction layer Predictor(·) to generate the predicted power load value for the next time step t+1. The calculation formula is as follows: Finally, by minimizing the predicted value Compared with the actual load value y i,t+1 The mean squared error loss function Loss is used to train the entire model end-to-end. The formula for the loss function is: Where N is the number of nodes in the batch, if the loss converges to a stable value, the power load prediction result corresponding to the node to be tested is output based on the optimal model.

Images