Cross-court distributed energy generation power joint prediction method based on deep learning

Abstract

The invention discloses a cross-court distributed energy generation power joint prediction method based on deep learning, and relates to the technical field of power system intelligent prediction. The method is used for solving the problems of weak space-time relevance and data islanding in distributed energy generation power prediction. The method comprises the following steps: firstly, constructing a dynamic graph structure based on geographical proximity and meteorological condition similarity between stations, and capturing spatial propagation characteristics of power fluctuation; secondly, a dynamic graph attention network is adopted to fuse the features of the adjacent transformer areas, and environment enhancement features containing a spatial dependency relationship are generated; then, minute-level fluctuation features and hour-level trend features are extracted through a multi-scale convolution kernel, and the feature fusion proportion is dynamically adjusted according to meteorological stability; and finally, network parameters are updated on the support set transformer area by using a meta learning mechanism, and meta learner configuration is optimized on the query set transformer area, so that rapid adaptation of a new transformer area is realized. According to the method, the cross-court generation power prediction precision and the model generalization capability are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent prediction of power systems, in particular to a cross-district distributed energy power generation power joint prediction method based on deep learning. BACKGROUND

[0002] With the wide access of distributed photovoltaic, wind power and other new energy in the distribution network, the district-level power system gradually presents the characteristics of multi-source, dispersion and dynamics. Influenced by many factors such as weather conditions, geographical location and user load behavior, the power generation of distributed energy has obvious time-varying and regional coupling, making it difficult to accurately predict the power generation of distributed energy, which is an important basis for maintaining stable operation and optimal scheduling of power grid. However, in the multi-district collaborative environment, power fluctuation is not only determined by the characteristics of a single district, but also affected by the coupling of adjacent regional weather field distribution differences and geographical topological structure. How to describe the spatial correlation and time sequence variation between districts at the same time is a key problem to realize high-precision power generation prediction.

[0003] Existing distributed energy power generation prediction methods are mostly based on single-district time series models or static feature correlation models. Time series models usually assume that the power change process is stationary, which is difficult to deal with weather mutations or non-synchronous characteristics of local regional power fluctuations; although the graph model based on static features can describe the geographical proximity relationship to a certain extent, it ignores the dynamic correlation changes caused by the evolution of weather conditions over time, and cannot reflect the transmission effect of cross-district power fluctuations. In addition, some methods use fixed weight mechanism in the process of spatial feature extraction and time sequence feature fusion, which lacks adaptive response to environmental changes, resulting in significant decline in prediction accuracy in complex scenarios. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a cross-district distributed energy power generation power joint prediction method based on deep learning, which solves the problems in the above background art.

[0005] To achieve the above object, the application is implemented by the following technical solutions: 1. A cross-substation distributed energy power generation power joint prediction method based on deep learning, comprising the following steps: S1. Obtain historical monitoring data of each substation in the target area, calculate dynamic correlation weights based on the geographical proximity and meteorological condition similarity between substations, construct a dynamic graph structure reflecting the correlation relationship of power fluctuation between substations, and extract the time sequence features of each substation power sequence through a sliding time window to generate node initial features; S2. Input the dynamic graph structure and node initial features into a dynamic graph attention network, calculate the attention coefficients between nodes through dynamic correlation weights and node features, weight and fuse the features of adjacent nodes based on the attention coefficients, and generate environment enhanced features containing spatial dependence relationship, wherein the substation features with high geographical proximity obtain a basic weight, and the substation features with high meteorological field similarity obtain an enhanced weight; S3. Concatenate the node initial features and the environment enhanced features, extract minute-level fluctuation features and hour-level trend features through different scale convolution kernels respectively, dynamically adjust the fusion ratio of the two types of features according to the stability index of real-time meteorological data, and output the power generation power prediction value of each substation in the future period; S4. Construct a meta-training task containing a support set and a query set, calculate the prediction loss on the support set substation and update the parameters of the dynamic graph attention network, and evaluate the prediction accuracy of the dynamic graph attention network on the query set substation and optimize the parameter configuration of the meta-learner.

[0006] Further, the logical process of constructing the dynamic graph structure reflecting the correlation relationship of power fluctuation between substations is as follows: obtain the geographic coordinate information and real-time meteorological monitoring data of each substation in the target area, obtain the geographical proximity matrix by calculating the spatial distance between substations, and obtain the meteorological similarity matrix by analyzing the correlation of meteorological elements; combine the geographical proximity matrix and the meteorological similarity matrix according to the preset fusion rule, wherein the weight coefficient of the meteorological similarity matrix is greater than that of the geographical proximity matrix, to generate a dynamic correlation weight matrix reflecting the correlation strength of power fluctuation between substations; based on the dynamic correlation weight matrix, construct the adjacency matrix of the graph structure, map each substation to a graph node, and take the dynamic correlation weight as the edge weight, to form a dynamic graph structure representing the power fluctuation propagation relationship between substations.

[0007] Further, the specific process of generating the node initial features by extracting the time sequence features of each substation power sequence through the sliding time window is as follows: a plurality of sliding windows with different time lengths are set to segmentally sample the power sequence, wherein the short time window is used to capture the rapid fluctuation feature, and the long time window is used to capture the trend change feature; multi-dimensional feature extraction is performed on the power sequence in each window segment, including time domain statistical features and differential change features, the time domain statistical features cover mean, variance and extreme value statistics, and the differential change features include first-order differential and second-order differential; the feature sequences extracted by the windows are spliced and combined, compressed feature representations are obtained through feature dimension reduction processing, and node initial features representing the operation characteristics of substations are generated.

[0008] Further, the specific process of inputting the dynamic graph structure and the node initial features into the dynamic graph attention network to calculate the attention coefficients between nodes through the dynamic correlation weight and the node features is as follows: the dynamic correlation weight matrix is used as the prior distribution of attention calculation, and the weighted fusion is performed on the inner product operation result of the node features; the node features are linearly transformed through a learnable weight matrix to obtain a query vector and a key vector, the dot product attention of the query vector and the key vector is calculated, the dynamic correlation weight is combined for softmax normalization, and the attention coefficient distribution between nodes is obtained.

[0009] Further, the specific process of generating the environment enhanced features containing the spatial dependency relationship by weighting and fusing the features of adjacent nodes based on the attention coefficients is as follows: important neighbor nodes are screened according to the attention coefficients, neighbor nodes with attention coefficients exceeding a threshold value are retained, and the neighbor node features screened out are aggregated through attention weighting, wherein the attention coefficient based on the geographical proximity is used as the basic weight, and the attention coefficient based on the meteorological similarity is used as the enhanced weight; the neighbor features after weighted aggregation are spliced with the current node features, and the environment enhanced features fusing the spatial context information are generated through a nonlinear transformation layer.

[0010] Further, the specific process of splicing the node initial features and the environment enhanced features and extracting the minute-level fluctuation features and the hour-level trend features through convolution kernels with different scales is as follows: the node initial features and the environment enhanced features are spliced along the feature dimension to form fusion features containing the operation characteristics of substations and the spatial correlation characteristics; dense convolution operation is performed on the fusion features along the time dimension through a narrow convolution kernel to extract high-frequency fluctuation features reflecting the minute-level power rapid change; convolution calculation is performed on the fusion features along the time dimension through a wide convolution kernel to extract low-frequency trend features reflecting the hour-level power slow change; the high-frequency fluctuation features and the low-frequency trend features are respectively subjected to time dimension pooling processing to retain key feature information and reduce the feature dimension.

[0011] Further, the specific process of dynamically adjusting the fusion ratio of the two types of features according to the stability index of real-time meteorological data and outputting the power generation prediction value of each substation in the future period is as follows: the variance and change rate of real-time meteorological data are calculated to obtain a meteorological fluctuation index representing the fluctuation degree of meteorological conditions, and the fusion weights of high-frequency fluctuation features and low-frequency trend features are adjusted according to the comparison result of the meteorological fluctuation index and the preset threshold, when the meteorological fluctuation index is higher than the threshold, the fusion weight of high-frequency fluctuation features is increased, and when the meteorological fluctuation index is lower than the threshold, the fusion weight of low-frequency trend features is increased, and the multi-scale features after weighted fusion are nonlinearly transformed through a multilayer perception machine to output the power generation prediction sequence of each substation in the future period.

[0012] Further, the specific process of calculating the prediction loss on the support set substation and updating the parameters of the dynamic graph attention network is as follows: the forward propagation calculation is run on the support set substation to obtain the difference between the power generation prediction value and the actual measured value, and the prediction loss is quantified by a mean square error function; the gradients of the loss function to the trainable parameters in the dynamic graph attention network, including the attention weight matrix and the feature transformation parameter, are calculated by the error back propagation algorithm; the parameters of the dynamic graph attention network are updated by the gradient descent algorithm to adjust the mapping relationship of the network.

[0013] Further, the specific process of evaluating the prediction accuracy of the dynamic graph attention network on the query set substation and optimizing the parameter configuration of the meta-learner is as follows: the prediction output of the dynamic graph attention network is calculated on the query set substation, and the mean absolute error and root mean square error of the predicted power and the actual power are calculated; the optimization objective function of the meta-learner is constructed based on the prediction error on the query set, and the hyperparameters of the meta-learner are updated by the second-order derivative calculation; the parameter updating process is iteratively executed to optimize the configuration of the meta-learner and improve the prediction performance of the dynamic graph attention network on new substations.

[0014] The present application has the following beneficial effects: (1) The cross-substation distributed energy power joint prediction method based on deep learning dynamically depicts the transmission characteristics of power fluctuations between different substations through the joint calculation of geographical proximity and meteorological condition similarity, overcoming the shortcomings of traditional static graph models that cannot reflect the influence of meteorological field changes on power correlation. Further, based on the adaptive weighting mechanism of the dynamic graph attention network, the model can dynamically adjust the fusion strength of neighborhood features under changing meteorological conditions, thereby improving the accuracy and stability of cross-substation power correlation modeling and realizing self-learning expression of complex spatial dependence relationships.

[0015] (2) The cross-district distributed energy power generation power joint prediction method based on deep learning can capture minute-level fluctuation patterns and hour-level trend changes by performing multi-scale convolution on the node initial features and environmental enhanced features, thereby improving the response ability of the prediction results to short-term disturbances and long-term trends. Combined with the meta-learning mechanism, the parameters are transferred and optimized between the support set and the query set, effectively improving the generalization performance and prediction accuracy of the model in different regions and different weather conditions, and realizing the joint intelligent prediction of the cross-district distributed energy power generation power.

[0016] Of course, implementing any product of the present application does not necessarily require all the advantages described above to be achieved at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the cross-district distributed energy power generation power joint prediction method based on deep learning of the present application. DETAILED DESCRIPTION

[0018] The cross-district distributed energy power generation power joint prediction method based on deep learning of the present application solves the problem that the existing prediction model is difficult to simultaneously depict the power space correlation between districts and the dynamic change of weather conditions, resulting in unstable prediction accuracy.

[0019] The scheme in the embodiments of the present application has the following overall idea: by establishing a dynamic graph structure that can reflect geographical proximity and meteorological similarity, the adaptive weighted fusion of features between districts is realized by combining the attention mechanism, the high-frequency fluctuation features and low-frequency trend features of the power sequence are extracted by using multi-scale convolution, the feature fusion ratio is dynamically adjusted according to the real-time meteorological stability, and finally the meta-learning mechanism is introduced to realize the transfer and optimization of model parameters between different districts, thereby improving the accuracy and cross-regional generalization ability of power generation power prediction.

[0020] Please refer to Figure 1This invention provides a technical solution: a deep learning-based method for joint prediction of distributed energy generation power across power distribution areas, comprising the following steps: S1. Obtaining historical monitoring data of each power distribution area within the target region, calculating dynamic correlation weights based on the geographical proximity and meteorological condition similarity between power distribution areas, constructing a dynamic graph structure reflecting the power fluctuation correlation between power distribution areas, and simultaneously extracting the temporal features of the power sequence of each power distribution area through a sliding time window to generate initial node features; S2. Inputting the dynamic graph structure and initial node features into a dynamic graph attention network, calculating the attention coefficients between nodes through the dynamic correlation weights and node features, and weighting and fusing the features of adjacent nodes based on the attention coefficients to generate a graph containing... The environment enhancement features are spatially dependent, with features of stations with high geographical proximity receiving basic weights and features of stations with high meteorological field similarity receiving enhanced weights; S3. The initial node features are concatenated with the environment enhancement features, and minute-level fluctuation features and hour-level trend features are extracted by convolutional kernels of different scales. The fusion ratio of the two types of features is dynamically adjusted according to the stability index of real-time meteorological data, and the predicted power generation value of each station for future periods is output; S4. A meta-training task containing a support set and a query set is constructed. The prediction loss is calculated and the parameters of the dynamic graph attention network are updated on the stations in the support set. The prediction accuracy of the dynamic graph attention network is evaluated and the parameter configuration of the meta-learner is optimized on the stations in the query set.

[0021] In this implementation plan, S1. In this step, the term "transformer area" refers to a power supply area unit centered on a transformer in a distribution network. Different transformer areas are geographically distributed, and their power generation is spatially correlated due to factors such as meteorological conditions and load characteristics. By collecting historical monitoring data (including power output, irradiance, and temperature) of each transformer area, the geographical proximity (i.e., a measure of spatial distance) between adjacent transformer areas and the similarity between meteorological conditions (which can be measured by methods such as the Pearson correlation coefficient) are calculated, thereby obtaining dynamic correlation weights reflecting the spatial and meteorological relationship. The dynamic graph structure constructed based on these weights is a graph network in which the connection relationships between nodes and edges are dynamically adjusted with changes in time or meteorological conditions, used to more realistically depict the propagation law of power fluctuations between transformer areas. By extracting local time-series segments of the power sequence through a sliding time window, short-term dynamic characteristics of power changes can be captured. The finally generated node initial features comprehensively reflect the basic power state and temporal characteristics of the transformer area, providing input for subsequent spatial feature fusion. This involves jointly modeling the spatial relationships and temporal features between power stations, establishing a high-dimensional representation foundation for subsequent dynamic feature fusion. S2. The Dynamic Graph Attention Network (GNN) is a deep learning structure based on a Graph Neural Network (GNN) that adaptively adjusts the information propagation intensity based on the dynamic weights between nodes. This network determines the importance of each node within its neighborhood by calculating the attention coefficient, allowing the model to focus on power stations that have a significant impact on the target node. Through this mechanism, features of spatially close power stations receive basic weights during fusion, representing the power transmission effect caused by their geographical proximity; while power stations with similar meteorological conditions receive enhanced weights, reflecting the synchronous power change trend caused by meteorological disturbances. By adaptively learning the spatial dependency structure between power stations through the attention mechanism, the model can accurately capture the spatial transmission patterns of power fluctuations, thereby improving the robustness and generalization ability of spatial feature extraction. S3. In this step, the initial node features are first concatenated with the environmental enhancement features along the feature dimension to obtain a fused feature that includes both the temporal information of the power station itself and spatial correlation information. Subsequently, convolution kernels of different scales were used for feature extraction. Narrow convolution kernels were used to capture rapid fluctuations at the minute level (high-frequency features), while wide convolution kernels were used to extract smooth trends at the hour level (low-frequency features). Based on the changes in real-time meteorological data (such as light intensity, temperature, wind speed, etc.), stability indicators (such as variance or rate of change) were calculated, and the fusion ratio of high- and low-frequency features was dynamically adjusted accordingly.When weather fluctuations are strong, the model relies more on high-frequency features to respond to sudden changes; when the weather is stable, it focuses on low-frequency trends to improve long-term prediction stability. Through multi-scale convolution and dynamic feature fusion mechanisms, the power prediction model adaptively captures features at different time scales, thereby improving the accuracy and stability of prediction results across multiple time periods. In step S4, meta-learning is introduced, dividing the substation data into a support set and a query set to construct a multi-task training framework. The model first learns on the support set, updating the parameters of the dynamic graph attention network by minimizing the error (e.g., mean squared error) between the predicted and actual power. Then, the model's generalization performance is evaluated on the query set, optimizing the hyperparameters of the meta-learner by calculating higher-order gradients, enabling the model to quickly adapt to new substations or new weather conditions. The meta-learning mechanism enables cross-regional transfer and rapid adjustment of model parameters, significantly improving the model's predictive adaptability and training efficiency in unknown substations or changing environments.

[0022] Specifically, the logical process of constructing a dynamic graph structure reflecting the correlation of power fluctuations between power distribution areas is as follows: Geographic coordinates and real-time meteorological monitoring data of each power distribution area within the target region are obtained; a geographic proximity matrix is ​​obtained by calculating the spatial distance between power distribution areas, and a meteorological similarity matrix is ​​obtained by analyzing the correlation of meteorological elements; the geographic proximity matrix and the meteorological similarity matrix are combined according to a preset fusion rule, where the weight coefficient of the meteorological similarity matrix is ​​greater than the weight coefficient of the geographic proximity matrix, generating a dynamic correlation weight matrix reflecting the strength of the correlation between power fluctuations between power distribution areas; an adjacency matrix of the graph structure is constructed based on the dynamic correlation weight matrix, mapping each power distribution area as a graph node, and the dynamic correlation weight as the edge weight, forming a dynamic graph structure characterizing the propagation relationship of power fluctuations between power distribution areas.

[0023] In this implementation plan, firstly, the geographic coordinates and real-time meteorological monitoring data of each transformer substation within the target area are acquired. The geographic coordinates reflect the relative spatial positions of the substations, while the meteorological monitoring data includes various meteorological elements such as light intensity, ambient temperature, humidity, and wind speed, reflecting external environmental factors affecting the power generation capacity of distributed energy resources. Secondly, the spatial distance between any two transformer substations is calculated to measure their geographical proximity. The spatial distance can be calculated using the substation coordinates, specifically as follows: ;in, Indicates the area Taiwan Spatial distance between them; Taiwan District latitude and longitude coordinates; Taiwan District The latitude and longitude coordinates are used. Based on the distance results, a geographic proximity matrix is ​​constructed. ,in: In the formula, This is the distance attenuation coefficient, used to control the intensity of the influence of spatial distance on proximity. By using an exponential attenuation form, adjacent monitoring stations can be assigned higher proximity weights, thus highlighting the spatial correlation between geographically close stations. Simultaneously, feature extraction and correlation calculations are performed on meteorological monitoring data. Let the monitoring station area be... and The meteorological feature vectors are respectively and Then the meteorological similarity coefficient between the two stations can be expressed as: ;in, Indicates the correlation between meteorological elements; Indicates the area Meteorological feature vectors; This represents the magnitude of the vector. This similarity reflects the synchronicity and environmental consistency of meteorological conditions across different weather stations. Then, the geographic proximity matrix and the meteorological similarity matrix are weighted and combined according to a preset fusion rule to generate a dynamic correlation weight matrix. , ;in, The meteorological similarity weight coefficient has a value range of (0.5, 1) to ensure that meteorological factors play a dominant role in the association modeling. For meteorological similarity; Geographic proximity. Weighting coefficient. This can be automatically determined through cross-validation, specifically by minimizing the coefficients of the prediction error on the validation set. Finally, based on the dynamic association weight matrix... Construct an adjacency matrix with a graph structure, treating each transformer area as a graph node and dynamically assigning weights. As edge weights, a dynamic graph structure is formed, which can adaptively update with changes in meteorological and geographical conditions. This enables accurate modeling of the dynamic relationships between multiple factors between stations, allowing the graph structure to not only reflect spatial topology but also capture the time-varying impact of meteorological changes on power coupling.

[0024] Specifically, the process of extracting the temporal features of the power sequence of each transformer area through a sliding time window to generate the initial node features is as follows: Multiple sliding windows of different time lengths are set to sample the power sequence in segments, with short time windows used to capture rapid fluctuation features and long time windows used to capture trend change features; multi-dimensional features are extracted from the power sequence within each window segment, including time-domain statistical features and difference change features. The time-domain statistical features cover mean, variance, and extreme value statistics, while the difference change features include first-order and second-order differences; the feature sequences extracted from each window are concatenated and combined, and a compressed feature representation is obtained through feature dimensionality reduction, generating the initial node features characterizing the operating characteristics of the transformer area.

[0025] In this implementation scheme, firstly, multiple sliding time windows of different lengths are set for the power time series of each transformer area. For example, short windows are used to capture rapid power fluctuations (such as minute-level changes), and long windows are used to identify slow trends (such as hour-level changes). Assuming the time series is... Window length is Then at time step The window segment at that time is: ;in, Indicates the first Window length at time The power subsequence; For the first The length of each time window. Multi-dimensional feature extraction is performed on the power sequence within each window segment. The extracted features include: time-domain statistical features: calculating the mean, variance, and extreme values ​​of the power sequence to reflect the central tendency and dispersion of power output; and difference variation features: calculating the first-order and second-order differences to reflect the dynamic smoothness and acceleration of power changes over time. The first-order difference is defined as: The second-order difference is defined as: ;in, Indicates the characteristic of power change rate; This represents the acceleration characteristic of power change. Feature sequences extracted from different windows are concatenated and combined to form a feature matrix. Then, dimensionality reduction is performed using principal component analysis (PCA) or an autoencoder network to obtain the compressed feature representation vector. This represents the initial node feature at that moment. By combining multi-scale time windows with differential feature extraction, the short-term fluctuations and long-term trend features of the power sequence are captured, and the operating characteristics of the transformer area are represented in a compact high-dimensional vector form, providing an accurate temporal input basis for the spatial feature fusion in the subsequent graph network.

[0026] Specifically, the process of inputting the dynamic graph structure and initial node features into the dynamic graph attention network, and calculating the attention coefficients between nodes through dynamic association weights and node features is as follows: The dynamic association weight matrix is ​​used as the prior distribution for attention calculation, and weighted and fused with the inner product operation result of the node features; the node features are linearly transformed through the learnable weight matrix to obtain the query vector and key vector, the dot product attention of the query vector and key vector is calculated, and softmax normalization is performed in combination with the dynamic association weights to obtain the distribution of attention coefficients between nodes.

[0027] In this implementation scheme, the dynamic correlation weight matrix is ​​denoted as... ,in Represents a node With nodes The dynamic correlation strength between nodes serves as prior information for the attention distribution. First, for each node's feature vector... Through learnable linear transformation matrices and Mapped to query vectors respectively With key vector Subsequently, the computing nodes... For nodes Attention response value: ;in, Represents a node For nodes Attention coefficient; This is a temperature coefficient used to control the smoothness of attention distribution; This represents the similarity score between the query and the key. Prior weights provided for the dynamic graph structure are used to enhance the connection strength between nodes with strong spatial or meteorological correlations. Through the above calculations, the attention coefficient can adaptively integrate the spatial dependence and meteorological commonalities between power stations, enabling the model to more accurately capture the dynamic interaction relationships across power generation areas when predicting power output. Temperature coefficient The optimal value can be determined through a validation set search method, i.e., iterative adjustment within a preset range. The parameter with the smallest prediction error is selected as the final configuration. This method allows the sparsity of the attention distribution to adaptively adjust with scene features, improving the model's generalization performance.

[0028] Specifically, the process of generating environmental enhancement features containing spatial dependencies by weighted fusion of features of neighboring nodes based on attention coefficients is as follows: Important neighboring nodes are selected based on attention coefficients, and neighboring nodes with attention coefficients exceeding a threshold are retained. The features of the selected neighboring nodes are then aggregated with attention weights, where attention coefficients based on geographical proximity are used as basic weights and attention coefficients based on meteorological similarity are used as enhancement weights. The weighted aggregated neighboring features are then concatenated with the current node features, and environmental enhancement features that fuse spatial context information are generated through a nonlinear transformation layer.

[0029] In this implementation scheme, the specific process of generating environment enhancement features containing spatial dependency clauses by weighted fusion of features of adjacent nodes based on the attention coefficient is as follows: First, based on the attention coefficient... Filter neighboring nodes, retaining those with an attention coefficient greater than a threshold. The node as an important neighbor set , where the threshold Based on the attention coefficient distribution Quantiles are determined to adaptively control the number of neighbors. For each node... Enhance its environmental features Calculated as ;in, Neighboring nodes eigenvectors; and These represent the attention coefficients calculated based on geographic proximity and meteorological similarity, respectively. The basic weighting coefficient reflects the importance of geographical proximity. To enhance the weighting coefficients and improve the influence of weather stations with similar meteorological conditions; For node bias terms; This is a nonlinear activation function used to enhance the nonlinearity of feature representation. Through the above process, the model can dynamically aggregate the features of key neighbor nodes according to the attention distribution, achieving a joint expression of spatially dependent information and environmental features. By introducing a dual weighting mechanism of geography and meteorology, the environmental enhancement features retain the local consistency of physical adjacency between stations while strengthening the global similarity driven by meteorology, thereby improving the accuracy and robustness of distributed energy power prediction.

[0030] Specifically, the process of concatenating the initial node features with the enhanced environment features and extracting minute-level fluctuation features and hour-level trend features using convolutional kernels of different scales is as follows: The initial node features and the enhanced environment features are concatenated along the feature dimension to form a fused feature that includes the operating characteristics and spatial correlation characteristics of the transformer substation itself; dense convolution operations are performed on the fused feature along the time dimension using narrow convolutional kernels to extract high-frequency fluctuation features reflecting rapid minute-level power changes; convolution operations are performed on the fused feature along the time dimension using wide convolutional kernels to extract low-frequency trend features reflecting slow hour-level power changes; pooling is performed on the high-frequency fluctuation features and low-frequency trend features along the time dimension to retain key feature information and reduce feature dimensionality.

[0031] In this implementation scheme, firstly, the initial feature vector of the node is... With the corresponding environment-enhanced feature vector The features are concatenated along the feature dimension to form a fused feature vector: ;in, This represents a vector concatenation operation; Let n be the initial characteristics of the nodes in the transformer area n; To enhance the characteristics of its environment; This represents a fused representation that simultaneously incorporates the operational characteristics of the transformer substation and its spatial dependence features. This concatenation process ensures that the model can consider both local dynamic information and the influence of the external environment in subsequent convolutional processing. Subsequently, to capture the power generation variation characteristics at different time scales, the fused feature sequence is processed... Temporal convolution is performed using two different sets of convolution kernels: ; ;in, It exhibits high-frequency fluctuation characteristics on the minute level; It exhibits low-frequency trend characteristics at the hourly level; , These are the weight coefficients for the short and long convolution kernels, respectively; This indicates that the time span of a short convolutional kernel is smaller than that of a long convolutional kernel; , For bias terms; The activation function is used to enhance the nonlinear expressive power of the features. High-frequency fluctuation features extracted by short convolutional kernels reflect rapid changes in distributed energy output, such as power generation fluctuations caused by short-term cloud cover or sudden changes in local wind speed. Low-frequency trend features extracted by long convolutional kernels capture the overall operating trend of the distribution area and stable changes driven by meteorological conditions. Finally, time-dimensional pooling operations retain significant feature responses and compress redundant information, thus providing multi-scale dynamic input for subsequent prediction stages. Parallel feature extraction using convolutional kernels of different scales effectively balances the prediction model's ability to perceive short-term sensitive fluctuations and long-term trend changes, improving the model's temporal modeling accuracy under complex meteorological conditions.

[0032] Specifically, the process of dynamically adjusting the fusion ratio of the two types of features based on the stability index of real-time meteorological data to output the predicted power generation value for each substation area in the future period is as follows: Calculate the variance and rate of change of the real-time meteorological data to obtain the meteorological fluctuation index, which characterizes the degree of fluctuation in meteorological conditions. Adjust the fusion weights of high-frequency fluctuation features and low-frequency trend features based on the comparison results between the meteorological fluctuation index and a preset threshold. When the meteorological fluctuation index is higher than the threshold, increase the fusion weight of high-frequency fluctuation features; when the meteorological fluctuation index is lower than the threshold, increase the fusion weight of low-frequency trend features. Perform nonlinear transformation on the weighted fused multi-scale features through a multilayer perceptron to output the predicted power generation sequence for each substation area in the future period.

[0033] In this implementation plan, firstly, the variance and rate of change of meteorological elements (including light intensity, wind speed, and temperature) at the current moment are calculated to measure the stability of meteorological conditions. Let the meteorological characteristic sequence be... Then the meteorological fluctuation index It can be represented as: ;in, The variance representing meteorological characteristics; Indicates the rate of change of weather conditions between adjacent time points; , This is a balance coefficient used to adjust the contribution ratio of the variance term to the variation term. (Meteorological fluctuation index) This reflects the overall instability of meteorological conditions. A higher value indicates a stronger impact of environmental factors on power generation fluctuations, and the model should pay more attention to short-term dynamic characteristics. This should be based on meteorological fluctuation indicators and preset thresholds. Based on the comparison results, the fusion weights of minute-level fluctuation characteristics and hourly-level trend characteristics are dynamically adjusted: ;in, Indicates time The fusion characteristics; and The fusion weights are respectively the minute-level and hour-level features, satisfying... ;when At that time, linear scaling was used. To enhance the influence of high-frequency characteristics; conversely, to improve This is done to highlight trend characteristics. Finally, the feature sequences will be merged. Input a multilayer perceptron model and perform nonlinear mapping to output the power generation prediction sequence for each transformer area in the future time period: ;in, The mapping function representing the multilayer perceptron; Taiwan District The predicted power sequence is obtained. By introducing a meteorological fluctuation index to dynamically adjust the feature fusion ratio, the sensitivity of short-term predictions can be enhanced when the weather is unstable, and the stability of long-term trends can be highlighted when the weather is stable. This allows the model to maintain high prediction accuracy and robustness under multiple climate and scenario conditions.

[0034] Specifically, the process of calculating the prediction loss and updating the parameters of the dynamic graph attention network on the support cluster area is as follows: Forward propagation is run on the support cluster area to obtain the difference between the predicted and actual power generation values, and the prediction loss is quantified using the mean square error function; the gradient of the loss function with respect to the trainable parameters in the dynamic graph attention network, including the attention weight matrix and feature transformation parameters, is calculated using the error backpropagation algorithm; the parameters of the dynamic graph attention network are updated using the gradient descent algorithm to adjust the network's mapping relationship.

[0035] In this implementation scheme, firstly, the power time series and corresponding meteorological characteristics from the data supporting the power distribution area are input into the aforementioned trained dynamic graph attention network, and forward propagation calculation is performed to obtain the predicted power output sequence. Let the actual observed power sequence be... The mean squared error (MSE) function is then used to measure the prediction error and calculate the prediction loss: ;in, This represents the predicted loss for the support set; To support the number of samples in the set; For the first The actual power of each sample; This represents the corresponding predicted power value. The loss function measures the mean squared difference between the model's predicted output and the actual power; a smaller value indicates a better fit. After obtaining the loss value, the gradient of the loss function with respect to each trainable parameter in the dynamic graph attention network, including the attention weight matrix, is calculated using the error backpropagation algorithm. and characteristic transformation matrix Key parameters, such as the attention weight matrix, can be used as an example to calculate its gradient: ;in, This represents the gradient of the loss function with respect to the predicted output; This represents the partial derivative of the predicted output with respect to the attention weight matrix. Subsequently, the parameters in the dynamic graph attention network are updated using the gradient descent algorithm: ;in, Indicates the first The parameter values ​​at the next iteration; The learning rate controls the update step size. The learning rate can be adaptively adjusted based on the loss decline trend, enabling the network to converge quickly in the early stages and maintain stable updates in the later stages. By calculating the predicted loss on the support set of transformer areas and updating the dynamic graph attention network parameters, the network achieves adaptive learning of typical transformer area data patterns. This allows the model to more accurately capture the dynamic power fluctuation patterns between multiple transformer areas, providing a stable feature base for subsequent cross-transformer area migration.

[0036] Specifically, the process of evaluating the prediction accuracy of the dynamic graph attention network and optimizing the parameter configuration of the meta-learner on the query set is as follows: Calculate the prediction output of the dynamic graph attention network on the query set, and statistically analyze the mean absolute error and root mean square error between the predicted power and the actual power; construct the optimization objective function of the meta-learner based on the prediction error on the query set, and update the hyperparameters of the meta-learner by calculating the second derivative; iteratively execute the parameter update process to optimize the configuration of the meta-learner and improve the prediction performance of the dynamic graph attention network on the new query set.

[0037] In this implementation scheme, after the parameter update of the support set stage is completed, a dynamic graph attention network is applied to the query set area data to calculate the predicted output sequence. With actual power sequence The difference. To comprehensively measure prediction accuracy, the mean absolute error (MAE) and root mean square error (RMSE) are calculated separately, and their calculation formulas are as follows: ;in, The number of samples in the query set; For the first The actual power value of each sample; These are the model's predicted values. MAE reflects the average magnitude of the model's prediction error, while RMSE emphasizes the penalty for large error samples. Combining the two provides a comprehensive evaluation of the model's predictive ability under different fluctuation ranges. The optimization objective function of the meta-learner is constructed based on the error metrics of the query set: ;in, Optimize the objective for meta-learning; , The balancing coefficient is used to adjust the contribution ratio of different error terms. The meta-learner uses a second-order gradient-based optimization strategy to update the hyperparameters, that is, by calculating the second derivative of the loss function with respect to the model parameters, it obtains the sensitivity of the meta-parameters to the model's generalization performance, and updates the meta-learner parameters accordingly. ;in, This represents the parameter set of the meta-learner; The meta-learning rate controls the step size for updating meta-parameters. Through multiple rounds of iterative optimization, the meta-learner can quickly adjust the initial parameters of the dynamic graph attention network when migrating between different power distribution areas, improving its prediction accuracy and generalization ability in new power distribution area scenarios. By evaluating the prediction accuracy of the dynamic graph attention network on the query set and using the meta-learning optimization strategy to adjust the parameters, the model can not only adapt to the power characteristics of known power distribution areas, but also maintain high prediction reliability and migration adaptability in new power distribution areas with different meteorological conditions and geographical structures, thereby achieving adaptive aggregation prediction of cross-regional distributed energy generation power.

[0038] In summary, this application has at least the following effects: This deep learning-based method for joint prediction of distributed energy generation power across distribution areas effectively characterizes the correlation and temporal features of power fluctuations between distribution areas by constructing a dynamic graph structure and extracting initial node features. It significantly enhances the model's ability to capture spatial dependencies by using a dynamic graph attention network combined with geographical proximity and meteorological similarity for weighted feature fusion. Furthermore, it extracts minute-level fluctuation features and hourly-level trend features through multi-scale convolutional kernels and dynamically adjusts the fusion weights based on real-time meteorological data, achieving accurate modeling of short-term rapid fluctuations and long-term trend changes. The introduction of a meta-learning mechanism enables the model to quickly adapt to the query set of distribution areas based on support set learning, achieving cross-distribution area migration prediction. Therefore, this method can improve the accuracy and robustness of power generation prediction in multi-distribution area and multi-meteorological scenarios, significantly enhance the scheduling and operation management efficiency of distributed energy, and provide reliable decision support for smart grid operation.

[0039] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0040] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, 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 apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxesFigure 1 A device that provides the functions specified in one or more boxes.

[0041] These 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 function 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 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0042] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0043] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based cross-substation distributed energy power generation power joint prediction method, characterized in that, The method comprises the following steps: S1. Obtain historical monitoring data of each substation in the target area, calculate dynamic correlation weights based on the geographical proximity and meteorological condition similarity between substations, construct a dynamic graph structure reflecting the correlation relationship of power fluctuations between substations, and extract time sequence features of power sequences of each substation through a sliding time window to generate initial node features; S2. Input the dynamic graph structure and initial node features into a dynamic graph attention network, calculate the attention coefficients between nodes through dynamic correlation weights and node features, and generate environment-enhanced features containing spatial dependency relationships by weighting and fusing the features of adjacent nodes based on the attention coefficients, wherein the features of substations with high geographical proximity obtain a basic weight, and the features of substations with high meteorological field similarity obtain an enhanced weight; S3. Concatenate the initial node features and the environment-enhanced features, extract minute-level fluctuation features and hour-level trend features through different scale convolution kernels, dynamically adjust the fusion ratio of the two types of features according to the stability index of real-time meteorological data, and output the power prediction value of each substation in the future period; S4. Construct a meta-training task containing a support set and a query set, calculate the prediction loss on the support set substations and update the parameters of the dynamic graph attention network, and evaluate the prediction accuracy of the dynamic graph attention network on the query set substations and optimize the parameter configuration of the meta-learner.

2. The deep learning-based cross-substation distributed energy power generation power joint prediction method according to claim 1, characterized in that: The logical process of constructing a dynamic graph structure reflecting the correlation relationship of power fluctuations between substations is as follows: Obtain the geographical coordinate information and real-time meteorological monitoring data of each substation in the target area, obtain the geographical proximity matrix by calculating the spatial distance between substations, and obtain the meteorological similarity matrix by analyzing the correlation of meteorological elements; Combine the geographical proximity matrix and the meteorological similarity matrix according to the preset fusion rule, wherein the weight coefficient of the meteorological similarity matrix is greater than that of the geographical proximity matrix, to generate a dynamic correlation weight matrix reflecting the correlation strength of power fluctuations between substations; Based on the dynamic correlation weight matrix, construct the adjacency matrix of the graph structure, map each substation to a graph node, and use the dynamic correlation weight as the edge weight to form a dynamic graph structure representing the power fluctuation propagation relationship between substations.

3. The deep learning-based cross-substation distributed energy power generation power joint prediction method according to claim 2, characterized in that: The specific process of extracting time sequence features of power sequences of each substation through a sliding time window to generate initial node features is as follows: Set multiple sliding windows with different time lengths to segment and sample the power sequence, wherein short time windows are used to capture rapid fluctuation features, and long time windows are used to capture trend change features; Extract multi-dimensional features of the power sequence in each window segment, including time domain statistical features and difference change features, wherein the time domain statistical features include mean, variance and extreme value statistics, and the difference change features include first-order difference and second-order difference; Concatenate the feature sequences extracted from each window, compress the features through feature dimension reduction processing to obtain compressed feature representations, and generate initial node features representing the operating characteristics of the substations.

4. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 1, characterized in that: The specific process of inputting the dynamic graph structure and initial node features into the dynamic graph attention network to calculate the attention coefficients between nodes is as follows: The dynamic correlation weight matrix is used as the prior distribution of attention calculation, and the inner product operation result of node features is weighted and fused; The node features are linearly transformed by a learnable weight matrix to obtain query vectors and key vectors, the dot product attention of query vectors and key vectors is calculated, and the dynamic correlation weight is used for softmax normalization to obtain the attention coefficient distribution between nodes.

5. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 4, characterized in that: The specific process of weighting and fusing the features of adjacent nodes based on the attention coefficient to generate environment-enhanced features containing spatial dependency is as follows: According to the attention coefficient, important neighbor nodes are screened, and neighbor nodes with attention coefficients exceeding a threshold value are retained. The features of the screened neighbor nodes are aggregated with attention weighting, wherein the attention coefficient based on geographical proximity is used as the basic weight, and the attention coefficient based on meteorological similarity is used as the enhanced weight; The neighbor features after weighted aggregation are spliced with the current node features to generate environment-enhanced features that fuse spatial context information through a nonlinear transformation layer.

6. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 1, characterized in that: The specific process of splicing the node initial features with the environment-enhanced features and extracting minute-level fluctuation features and hour-level trend features through different scale convolution kernels is as follows: The node initial features and the environment-enhanced features are spliced along the feature dimension to form fused features containing the running characteristics of the transformer area itself and the spatial correlation characteristics; Dense convolution operation is performed on the fused features along the time dimension through a narrow convolution kernel to extract high-frequency fluctuation features reflecting the rapid change of minute-level power; Convolution calculation is performed on the fused features along the time dimension through a wide convolution kernel to extract low-frequency trend features reflecting the slow change of hour-level power; The high-frequency fluctuation features and the low-frequency trend features are respectively processed by time dimension pooling to retain key feature information and reduce feature dimension.

7. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 6, characterized in that: The specific process of dynamically adjusting the fusion ratio of the two types of features according to the stability index of real-time meteorological data to output the power prediction value of each transformer area in the future period is as follows: The variance and change rate of real-time meteorological data are calculated to obtain the meteorological fluctuation index representing the fluctuation degree of meteorological conditions. The fusion weights of high-frequency fluctuation features and low-frequency trend features are adjusted according to the comparison result of the meteorological fluctuation index and the preset threshold. When the meteorological fluctuation index is higher than the threshold, the fusion weight of high-frequency fluctuation features is increased; When the meteorological fluctuation index is lower than the threshold, the fusion weight of low-frequency trend features is increased; The multi-scale features after weighted fusion are nonlinearly transformed by a multilayer perceptron to output the power prediction sequence of each transformer area in the future period.

8. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 1, characterized in that: The specific process of calculating the prediction loss on the support set transformer area and updating the parameters of the dynamic graph attention network is as follows: Forward propagation calculation is performed on the support set transformer area to obtain the difference between the power prediction value and the actual measured value, and the prediction loss is quantified by the mean square error function; The gradients of the loss function with respect to the trainable parameters in the dynamic graph attention network, including the attention weight matrix and the feature transformation parameters, are calculated by the error back propagation algorithm; The parameters of the dynamic graph attention network are updated by the gradient descent algorithm to adjust the mapping relationship of the network.

9. The deep learning-based cross-district distributed energy power generation power joint prediction method according to claim 1, characterized in that: The specific process of evaluating the prediction accuracy of the dynamic graph attention network on the query set transformer area and optimizing the parameter configuration of the meta-learner is as follows: The prediction output of the dynamic graph attention network is calculated on the query set station area, and the mean absolute error and root mean square error of the predicted power and the actual power are counted; Based on the prediction error on the query set, the optimization objective function of the meta-learner is constructed, and the hyperparameters of the meta-learner are updated by calculating the second derivative; Iterative parameter updating process is performed to optimize the configuration of the meta-learner and improve the prediction performance of the dynamic graph attention network in the new station area.

Images