Park power grid flexible load resource optimization method and system based on deep learning

Abstract

The invention discloses a park power grid flexible load resource optimization method and system based on deep learning. The method comprises the steps of collecting operation data and meteorological information of each flexible load main body in a target park; preprocessing the operation data and the meteorological information to obtain multi-element load characteristics; according to the multi-element load characteristics, carrying out dynamic coupling on the association relationship between the weather and the electricity price fluctuation to obtain coupling characteristics; according to the multi-element load characteristics and the coupling characteristics, load prediction is carried out through a park load prediction model, and a park load prediction curve in a future set time window is obtained; according to the park load prediction curve, load resource adjustment is carried out on the flexible load main body in the target park; according to the method, the error of the park flexible load resource prediction result can be effectively reduced, the precision and robustness of park flexible load resource prediction are improved, and the accuracy and precision requirements of park flexible load resource scheduling and active power distribution management are met.

Description

Technical Field

[0001] The present invention relates to the field of flexible power grid technology, and in particular to a method and system for optimizing flexible load resources in a campus power grid based on deep learning. Background Art

[0002] As campus power grids transform toward intelligence and flexibility, refined forecasting of flexible load resources has become a key step in achieving load regulation, demand response, and coordinated supply and demand control. With a high proportion of new energy access and the coexistence of multiple energy-consuming terminals, the load characteristics of campus power grids are highly dynamic and nonlinear. Especially under extreme operating conditions such as high temperature, high humidity, or fluctuating electricity prices, the response behavior of flexible load resources such as air conditioners, electric vehicle charging stations, and interruptible industrial loads is extremely complex. Existing load forecasting methods are mostly based on single data sources, fixed features, and static models, making it difficult to capture the coupling relationships and spatiotemporal correlation patterns between different types of equipment. Furthermore, existing solutions lack systematic processing for the alignment and fusion of multi-source heterogeneous data, ignoring factors such as sampling rate differences, inconsistent spatial granularity, and load response lags. This results in large errors and poor robustness in load forecasting results, making it difficult to meet the accuracy and precision requirements of campus flexible load resource scheduling and active power distribution management. Summary of the Invention

[0003] In response to the problems existing in the existing technology, an embodiment of the present invention provides a method and system for optimizing the flexible load resources of a campus power grid based on deep learning, which can effectively reduce the error of the flexible load resource prediction results of the campus, improve the accuracy and robustness of the flexible load resource prediction of the campus, and meet the accuracy and precision requirements of the flexible load resource scheduling and active power distribution management of the campus.

[0004] In a first aspect, an embodiment of the present invention provides a method for optimizing flexible load resources in a campus power grid based on deep learning, comprising: Collecting operating data of each flexible load entity in the target park and meteorological information monitored by meteorological stations around the target park; wherein the operating data includes load information and electricity price information of the corresponding flexible load entity; Preprocessing the operating data and the meteorological information to obtain a multi-element load characteristic; wherein the multi-element load characteristic includes a load characteristic, a meteorological characteristic, and an electricity price characteristic; According to the multi-load characteristics, the correlation between the weather and the electricity price fluctuation is dynamically coupled to obtain the coupling characteristics; Based on the multi-element load characteristics and the coupling characteristics, load forecasting is performed using a pre-trained park load forecasting model to obtain a park load forecast curve within a future set time window; According to the park load forecast curve, load resources of the flexible load entities in the target park are adjusted.

[0005] As an improvement to the above solution, the operation data and the meteorological information are preprocessed to obtain multivariate load characteristics, including: Calculating the contribution of each flexible load subject according to the absolute deviation between the first historical average power consumption of each flexible load subject and the second average power consumption of the target park and a preset weight distribution steepness adjustment parameter; Performing time alignment on the load information in the operation data of each of the flexible load bodies using a dynamic time warping function to obtain the load characteristics of each of the flexible load bodies; Calculating the load characteristics of the target park according to the load characteristics and contribution of each of the flexible load entities; Performing spatial interpolation diffusion on the meteorological information using a preset convolution kernel to obtain meteorological characteristics of the meteorological information after spatial interpolation within the target park; Performing time alignment and interpolation processing on the electricity price information and the load time series characteristics to obtain the electricity price characteristics; The load characteristics, the meteorological characteristics and the electricity price characteristics of the target park are combined to obtain the multivariate load characteristics.

[0006] As an improvement to the above solution, the correlation between weather conditions and electricity price fluctuations is dynamically coupled based on the multi-load characteristics to obtain coupling characteristics, including: Calculating a mixed second-order partial derivative of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating the gradient Euclidean norm of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating a meteorological coupling weight based on a nonlinear activation function according to the mixed second-order partial derivative, the gradient Euclidean norm, and a preset sensitivity coefficient; Calculating the product of the temperature element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a temperature-electricity price product feature; Calculating the product of the humidity element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a humidity-electricity price product feature; The coupling feature is generated according to the meteorological coupling weight, the temperature electricity price product feature, and the humidity electricity price product feature.

[0007] As an improvement to the above solution, the training process of the park load forecasting model includes: Collecting historical operating data of each flexible load entity in the target park and historical meteorological information monitored by meteorological stations around the target park; wherein the historical operating data includes historical load information and historical electricity price information of the corresponding flexible load entity; Preprocessing the historical operating data and the historical meteorological information to obtain historical multi-element load characteristics; wherein the historical multi-element load characteristics include historical load characteristics, historical meteorological characteristics, and historical electricity price characteristics; According to the historical multi-load characteristics, the correlation between meteorological conditions and electricity price fluctuations is dynamically coupled to obtain historical coupling characteristics; According to the historical multi-load characteristics and the historical coupling characteristics, the park load forecasting model is trained until a preset convergence condition is met, thereby obtaining a trained park load forecasting model.

[0008] As an improvement to the above solution, the park load forecasting model is trained according to the historical multi-load characteristics and the historical coupling characteristics until a preset convergence condition is met, thereby obtaining a trained park load forecasting model, including: splicing the historical multivariate load characteristics and the historical coupling characteristics to obtain an input characteristic matrix; Input the input feature matrix into the park load forecasting model built based on bidirectional temporal attention for model training to obtain the historical park forecast load curve output in this round; Calculate the loss value and KL divergence distribution matching fluctuation between the historical park forecast load curve and the corresponding park real load curve through a preset double penalty loss function; Determining whether the convergence condition is satisfied based on the loss value and the KL divergence distribution matching fluctuation; wherein the convergence condition includes: the decrease in the loss value for multiple consecutive rounds is less than a preset attenuation threshold, and the KL divergence distribution matching fluctuation for multiple consecutive rounds is less than a preset fluctuation threshold; If so, stop training the park load forecasting model and output the trained park load forecasting model; If not, continue training the park load forecasting model.

[0009] As an improvement to the above solution, the park load forecasting model includes an input layer arranged according to a hierarchical encoder architecture, a bidirectional temporal attention block, a feature distillation layer, a decoding layer built based on a gated deconvolution module, and an output layer; The forward temporal dependency features of the input feature matrix received by the input layer are captured by the causal convolution layer in the bidirectional temporal attention block to obtain a query vector matrix; Capturing backward historical pattern features of the input feature matrix received by the input layer through the time-reversed convolution layer in the bidirectional temporal attention block to obtain a key vector matrix and a value vector matrix; In the attention mechanism layer of the bidirectional temporal attention block, attention weights are calculated according to the query vector matrix and the key vector matrix, and an attention mechanism output matrix is ​​generated according to the attention weights, the value vector matrix, and the query vector matrix; The attention mechanism output matrix is ​​processed in sequence by the feature distillation layer, the decoding layer, and the output layer to generate a corresponding historical park predicted load curve.

[0010] As an improvement to the above solution, the method further includes: Performing a k-means cluster analysis on the input feature matrix to divide the input feature matrix into a plurality of clusters, wherein the cluster center of each cluster corresponds to a load mode; Calculate the variance of each cluster, and obtain the weight coefficient of the corresponding cluster according to the proportion of the variance of each cluster to the total variance of all clusters. Performing a principal component analysis dimensionality reduction operation on each of the cluster centers so that the dimensions of each cluster center meet the input requirements of the park load forecasting model; According to the weight coefficients of the clusters, weighted summation is performed on the cluster centers after dimensionality reduction to obtain an initialization weight matrix; The initialization weight matrix is ​​used to initialize the model parameters of the park load forecasting model.

[0011] As an improvement to the above solution, the dual-penalty loss function includes a mean absolute error term, a curvature constraint term, and a distribution matching term; The calculation of the loss value and KL divergence distribution matching fluctuation between the historical park predicted load curve and the corresponding park actual load curve using a preset double penalty loss function includes: Calculating the average of the absolute errors between the predicted load in the historical park predicted load curve and the actual load in the corresponding park actual load curve to obtain a mean absolute error term; Calculating the second-order difference of the predicted load in the historical park predicted load curve, and calculating the Frobenius norm of the second-order difference to obtain a curvature constraint term; Calculating the KL divergence distribution matching fluctuation between the probability distribution of the historical park predicted load curve and the probability distribution of the corresponding park actual load curve as a distribution matching item; According to the mean absolute error term, the curvature constraint term and the distribution matching term, a loss value between the historical park predicted load curve and the corresponding park actual load curve is obtained.

[0012] As an improvement to the above solution, the method further includes: At the beginning of each round of training, the decay rate of the learning rate of the current training round is calculated by the natural exponential function based on the ratio of the current training round to the preset total number of training rounds and the variance of the loss value of the current training round; The decay rate is adopted to dynamically adjust the learning rate of the park load forecasting model.

[0013] In a second aspect, an embodiment of the present invention provides a campus power grid flexible load resource optimization system based on deep learning, including: A data acquisition module, configured to collect operating data of each flexible load entity within a target park and meteorological information monitored by meteorological stations surrounding the target park; wherein the operating data includes load information and electricity price information of the corresponding flexible load entity; a data preprocessing module, configured to preprocess the operating data and the meteorological information to obtain a multivariate load characteristic; wherein the multivariate load characteristic includes the load characteristic, meteorological characteristic, and electricity price characteristic of each of the flexible load entities after time and space alignment; A feature coupling module is used to dynamically couple the correlation between meteorological conditions and electricity price fluctuations according to the multi-load characteristics to obtain coupling features; A load forecasting module, configured to perform load forecasting using a pre-trained park load forecasting model based on the multi-element load characteristics and the coupling characteristics, and obtain a park load forecast curve within a future set time window; The load resource adjustment module is used to adjust the load resources of the flexible load body in the target park according to the park load forecast curve.

[0014] Compared with the prior art, the embodiment of the present invention provides a method and system for optimizing flexible load resources of a park power grid based on deep learning, which collects the operating data of each flexible load entity in the target park and the meteorological information monitored by the meteorological stations around the target park; wherein the operating data includes the load information and electricity price information of the corresponding flexible load entity; the operating data and the meteorological information are preprocessed to obtain multivariate load characteristics; wherein the multivariate load characteristics include load characteristics, meteorological characteristics and electricity price characteristics; according to the multivariate load characteristics, the correlation between meteorological and electricity price fluctuations is dynamically coupled to obtain coupling characteristics; according to the multivariate load characteristics and the coupling characteristics, load prediction is performed through a pre-trained park load prediction model to obtain a park load prediction curve within a future set time window; according to the park load prediction curve, load resources of the flexible load entities in the target park are adjusted; the present invention can effectively reduce the error of the park flexible load resource prediction results, improve the accuracy and robustness of the park flexible load resource prediction, and meet the accuracy and precision requirements of the park flexible load resource scheduling and active power distribution management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings used in the implementation methods. Obviously, the drawings described below are only some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 This is a flow chart of a method for optimizing flexible load resources in a campus power grid based on deep learning, provided by an embodiment of the present invention; Figure 2 1 is a schematic diagram comparing curves of absolute errors obtained by using different prediction methods provided by an embodiment of the present invention; Figure 3 is a comparative schematic diagram of load distribution using different loss functions provided by an embodiment of the present invention; Figure 4 24-hour mean absolute error curve comparison diagram obtained by using different prediction methods provided by the embodiment of the present invention; Figure 5 1 is a schematic diagram of the scatter distribution of prediction results obtained by using ARIMA for load forecasting according to an embodiment of the present invention; Figure 6 Schematic diagram of the scatter distribution of prediction results obtained by using LSTM for load forecasting according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the scatter distribution of prediction results obtained by using GNN-GRU for load forecasting provided by an embodiment of the present invention; Figure 8 1 is a schematic diagram of the scattered distribution of prediction results obtained by using the embodiment of the present invention to perform load forecasting; Figure 9 This is a structural block diagram of a campus power grid flexible load resource optimization system based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] It is understood that the various numbers used in the embodiments of the present invention are merely for ease of description and are not intended to limit the scope of this application. The order of execution of each process does not necessarily imply a specific order of execution. The order of execution of each process should be determined by its function and inherent logic.

[0019] In the embodiments of the present invention, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. The terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, the elements defined by the sentence "includes..." do not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the elements.

[0020] See also Figure 1 , Figure 1 This is a flow chart of a method for optimizing flexible load resources in a campus power grid based on deep learning provided by an embodiment of the present invention. The method for optimizing flexible load resources in a campus power grid based on deep learning specifically includes: S11: Collecting operating data of each flexible load entity in the target park and meteorological information monitored by meteorological stations around the target park; wherein the operating data includes load information and electricity price information of the corresponding flexible load entity; S12: Preprocessing the operating data and the meteorological information to obtain a multi-element load characteristic; wherein the multi-element load characteristic includes a load characteristic, a meteorological characteristic, and an electricity price characteristic; S13: dynamically coupling the correlation between weather conditions and electricity price fluctuations based on the multi-load characteristics to obtain coupling characteristics; S14: performing load forecasting using a pre-trained park load forecasting model based on the multi-load characteristics and the coupling characteristics to obtain a park load forecast curve within a future set time window; S15: According to the park load forecast curve, load resources are adjusted for the flexible load entities in the target park.

[0021] In an embodiment of the present invention, the current operating data of each flexible load entity in the target park is collected, including timestamp, equipment identification of the corresponding flexible load entity, load information, electricity price information, etc.; the meteorological information monitored by the meteorological stations around the target park includes timestamp, temperature, humidity, etc. Then, the collected operating data and meteorological information are preprocessed to construct a multivariate load feature; a multivariate load feature is formed that includes the load characteristics (load information), meteorological characteristics (temperature and humidity), and electricity price characteristics (electricity price information) of each flexible load entity; based on the multivariate load feature, the correlation between meteorological and electricity price fluctuations is dynamically coupled to obtain a coupling feature; then the multivariate load feature and the coupling feature are spliced ​​and input into a pre-trained park load forecasting model for load forecasting to obtain a park load forecast curve within a future set time window; finally, based on the park load forecast curve, the load resources of the flexible load entities in the target park are adjusted; by dynamically coupling the correlation between meteorological and electricity price fluctuations, the present invention can explore the impact of the interaction between meteorological and electricity price fluctuations on the load demand of the flexible load entities in the target park, thereby reducing the error of the model's prediction results for the park's flexible load resources, improving the accuracy and robustness of the park's flexible load resource prediction, and meeting the accuracy and precision requirements of the park's flexible load resource scheduling and active power distribution management.

[0022] In an optional embodiment, the training process of the park load forecasting model includes: Collecting historical operating data of each flexible load entity in the target park and historical meteorological information monitored by meteorological stations around the target park; wherein the historical operating data includes historical load information and historical electricity price information of the corresponding flexible load entity; Preprocessing the historical operating data and the historical meteorological information to obtain historical multi-element load characteristics; wherein the historical multi-element load characteristics include historical load characteristics, historical meteorological characteristics, and historical electricity price characteristics; According to the historical multi-load characteristics, the correlation between meteorological conditions and electricity price fluctuations is dynamically coupled to obtain historical coupling characteristics; According to the historical multi-load characteristics and the historical coupling characteristics, the park load forecasting model is trained until a preset convergence condition is met, thereby obtaining a trained park load forecasting model.

[0023] For example, to predict the demand for flexible load resources within a target park, smart meters and IoT sensors deployed at distribution network nodes collect minute-by-minute operational data from each flexible load entity (i.e., electrical equipment such as air conditioning systems, electric vehicle charging stations, and interruptible industrial equipment). This data is then used to simultaneously access meteorological information from nearby meteorological stations within the target park, provided by the Meteorological Bureau, and electricity price information published by the power trading platform. Historical operational data, meteorological information, and electricity price information within a set time period are collected to construct a raw data set containing timestamps, device identifiers, historical load information, meteorological information, and electricity price information.

[0024] Furthermore, based on the control instructions of the flexible load entities recorded by the power grid dispatching system, the collected original data sets are finely labeled through manual labeling. The specific labeling process is as follows: First, based on the effective period of the control instructions, the "response period" and "baseline period" are marked in the historical load information (load time series) of each flexible load entity. The "response period" indicates the effective period of the control instructions, and the "baseline period" indicates the effective period of the non-control instructions.

[0025] Then, based on the device identification, query the preset device archive to find the type attributes and adjustable capability parameters (such as maximum interruption duration and power adjustment range) of each flexible load entity, and mark the type attributes, such as temperature control type, energy storage type, and production type; at the same time, mark the adjustable capability parameters, such as maximum interruption duration and power adjustment range.

[0026] Finally, based on historical meteorological information and historical electricity price information, find out meteorological warning events (such as extreme temperatures, for example, temperatures exceeding the set temperature range are considered extreme temperatures) and electricity price fluctuation events (such as electricity prices corresponding to peak and valley periods, which can be preset or the periods within a day corresponding to when the park load exceeds the set load threshold), and mark abnormal operating condition labels for meteorological warning events and electricity price fluctuation events.

[0027] In an optional embodiment, the training of the park load forecasting model according to the historical multi-load characteristics and the historical coupling characteristics until a preset convergence condition is met to obtain a trained park load forecasting model includes: splicing the historical multivariate load characteristics and the historical coupling characteristics to obtain an input characteristic matrix; Input the input feature matrix into the park load forecasting model built based on bidirectional temporal attention for model training to obtain the historical park forecast load curve output in this round; Calculate the loss value and KL divergence distribution matching fluctuation between the historical park forecast load curve and the corresponding park real load curve through a preset double penalty loss function; Determining whether the convergence condition is satisfied based on the loss value and the KL divergence distribution matching fluctuation; wherein the convergence condition includes: the decrease in the loss value for multiple consecutive rounds is less than a preset attenuation threshold, and the KL divergence distribution matching fluctuation for multiple consecutive rounds is less than a preset fluctuation threshold; If so, stop training the park load forecasting model and output the trained park load forecasting model; If not, continue training the park load forecasting model.

[0028] The park load forecasting model includes an input layer, a bidirectional temporal attention block, a feature distillation layer, a decoding layer based on a gated deconvolution module, and an output layer arranged according to a hierarchical encoder architecture. The forward temporal dependency features of the input feature matrix received by the input layer are captured by the causal convolution layer in the bidirectional temporal attention block to obtain a query vector matrix; Capturing backward historical pattern features of the input feature matrix received by the input layer through the time-reversed convolution layer in the bidirectional temporal attention block to obtain a key vector matrix and a value vector matrix; In the attention mechanism layer of the bidirectional temporal attention block, attention weights are calculated according to the query vector matrix and the key vector matrix, and an attention mechanism output matrix is ​​generated according to the attention weights, the value vector matrix, and the query vector matrix; The attention mechanism output matrix is ​​processed in sequence by the feature distillation layer, the decoding layer, and the output layer to generate a corresponding historical park predicted load curve.

[0029] Furthermore, the method further comprises: Performing a k-means cluster analysis on the input feature matrix to divide the input feature matrix into a plurality of clusters, wherein the cluster center of each cluster corresponds to a load mode; Calculate the variance of each cluster, and obtain the weight coefficient of the corresponding cluster according to the proportion of the variance of each cluster to the total variance of all clusters. Performing a principal component analysis dimensionality reduction operation on each of the cluster centers so that the dimensions of each cluster center meet the input requirements of the park load forecasting model; According to the weight coefficients of the clusters, weighted summation is performed on the cluster centers after dimensionality reduction to obtain an initialization weight matrix; The initialization weight matrix is ​​used to initialize the model parameters of the park load forecasting model.

[0030] The double penalty loss function includes a mean absolute error term, a curvature constraint term, and a distribution matching term. The calculation of the loss value and KL divergence distribution matching fluctuation between the historical park predicted load curve and the corresponding park actual load curve using a preset double penalty loss function includes: Calculating the average of the absolute errors between the predicted load in the historical park predicted load curve and the actual load in the corresponding park actual load curve to obtain a mean absolute error term; Calculating the second-order difference of the predicted load in the historical park predicted load curve, and calculating the Frobenius norm of the second-order difference to obtain a curvature constraint term; Calculating the KL divergence distribution matching fluctuation between the probability distribution of the historical park predicted load curve and the probability distribution of the corresponding park actual load curve as a distribution matching item; According to the mean absolute error term, the curvature constraint term and the distribution matching term, a loss value between the historical park predicted load curve and the corresponding park actual load curve is obtained.

[0031] Furthermore, the method further comprises: At the beginning of each round of training, the decay rate of the learning rate of the current training round is calculated by the natural exponential function based on the ratio of the current training round to the preset total number of training rounds and the variance of the loss value of the current training round; The decay rate is adopted to dynamically adjust the learning rate of the park load forecasting model.

[0032] The embodiment of the present invention takes into account the multi-source heterogeneous characteristics of the target park power grid load information, the large differences in the load information sampling rates of different flexible load entities, the meteorological information usually only has a single point of monitoring but needs to represent the entire park space, and the electricity price information is a global single scalar. Using simple interpolation or downsampling methods for processing will destroy the temporal correlation between the load information of different flexible load entities, and ignore the spatial differences in meteorological conditions in different areas of the target park, resulting in serious loss of feature information after fusion. Based on this, the original data set after the above-mentioned annotation is preprocessed by the method of spatiotemporal dynamic alignment in the embodiment of the present invention. The specific preprocessing process is as follows: First, the absolute deviation between the first historical average power consumption of the flexible load subject and the second historical average power consumption of the target park is calculated. Then, combined with the preset weight distribution steepness adjustment parameter, the energy consumption weight of each flexible load subject is dynamically calculated as the contribution of different flexible load subjects in feature fusion. The calculation formula of the energy consumption weight is as follows: (1); in, represents the energy consumption weight of the jth flexible load entity, represents the first historical average power consumption of the jth flexible load subject, characterizing the energy consumption level of the equipment, and M represents the number of flexible load subjects in the target park; is the natural exponential function; Indicates the second historical average power consumption of the target park, representing the overall energy consumption benchmark; represents the weight distribution steepness adjustment parameter, which controls the concentration of weight distribution, for example .

[0033] Then, the dynamic time warping function is applied, and the historical load information of different sampling rates of each flexible load body is time-aligned with a unified benchmark sampling rate as a reference, so as to retain the temporal correlation between different historical load information; at the same time, the spatial coordinates of the meteorological station are spatially convolved using a preset convolution kernel to achieve spatial interpolation and diffusion of meteorological information in the entire target source, making up for the shortcomings of single-point monitoring.

[0034] Afterwards, the load characteristics obtained after time and space alignment and the meteorological characteristics obtained after spatial diffusion are combined with the electricity price information at the corresponding time to ensure that the characteristics from different sources, different sampling rates, and different spatial granularities are consistent and integrated in the time and space dimensions, forming a feature vector after time and space alignment (i.e., historical multi-dimensional load characteristics) to avoid information loss. The function expression of the historical multi-dimensional load characteristics is as follows: (2); in, Represents the historical multivariate load characteristics after time-space alignment; for example, for the original data set, the data is divided according to the preset unit time, and the data corresponding to one time unit is a sample. Correspondingly, after the above preprocessing, the corresponding historical multivariate load characteristics can be obtained for one sample; represents the dynamic time warping function; Indicates the jth flexible load body from time t to t+ Historical load information between moments (i.e. load sequence), is the time window length (also called unit time), such as Hour; Represents a sequence of base sampling rates, with a unified sampling base, such as minute level; Represents the feature concatenation operator, which is used to connect different feature components; represents the convolution kernel (also called the meteorological diffusion convolution kernel), such as , used for spatial interpolation; represents the spatial convolution operator; represents the spatial coordinates of the meteorological station, and Represents the longitude and latitude dimensions respectively; express The electricity price at the time; Indicates the number of flexible load entities in the target park.

[0035] The generation process of the coupling characteristics of meteorological conditions and electricity prices is as follows: There is a dynamic nonlinear coupling effect between temperature and humidity meteorological factors and electricity price fluctuations. For example, the simultaneous occurrence of high temperature and high humidity may cause a surge in air conditioning load and push up electricity price peaks. If meteorological information and electricity price information are processed independently and this dynamic interaction mechanism between the two is ignored, it will be impossible to accurately capture the intensity and pattern of the impact of changing meteorological conditions on electricity prices. Based on this, the embodiment of the present invention quantifies the correlation strength between meteorology and electricity prices through coupling influencing factors and generates fusion features. First, the mixed second-order partial derivatives of the historical electricity price information with respect to the temperature and humidity in the historical meteorological information are calculated to reflect the intensity of the interaction between temperature and humidity on the change in electricity price at a specific moment. At the same time, the Euclidean L2 norm of the temperature and humidity gradient vectors in the historical meteorological information (i.e., the gradient L2 norm) is calculated to characterize the overall sensitivity of electricity prices to changes in meteorological factors. Then, the mixed second-order partial derivatives are adjusted by the sensitivity coefficient and input into a nonlinear activation function (such as the Sigmoid activation function). The mixed second-order partial derivatives are then multiplied by the gradient L2 norm to obtain the meteorological coupling weights at each moment (also called temperature-humidity coupling weights), which are used to dynamically adjust the contribution ratio of the temperature and humidity to the electricity price. The meteorological coupling weights are then used to perform a weighted fusion of the element-by-element product features of the temperature in the historical meteorological information and the electricity price in the historical electricity price information (i.e., the historical temperature-electricity-price product features) and the element-by-element product features of the humidity in the historical meteorological information and the electricity price in the historical electricity price information (i.e., the historical humidity-electricity-price product features). The final coupling features are formed through feature concatenation to dynamically capture the nonlinear interaction effect between meteorology and electricity prices. The specific expression is as follows: (3); (4); in, express Meteorological coupling weight at the moment; Represents the Sigmoid activation function, which is used to compress the input to interval, ensuring weight interpretability; represents the sensitivity coefficient, which controls the sensitivity of the coupling strength to the second-order derivative, such as ; express The mixed second-order partial derivative of the electricity price at the moment with respect to temperature and humidity, represents the temperature at time t, Represents the humidity at time t, which is calculated by fitting historical temperature and humidity data to reflect the impact of the interaction between temperature and humidity on electricity prices; Indicates the gradient of electricity price to temperature and humidity norm, is the gradient operator, which means to find the partial derivative of temperature T and humidity H. is the Euclidean L2 norm, which represents the overall sensitivity of electricity prices to meteorological factors; Represents historical coupling characteristics; represents the Hadamard product, element-wise product, in and Similarly, by performing the above processing on a sample of a unit time, the historical coupling characteristics corresponding to the sample can be obtained.

[0036] In some embodiments, a scatter plot in three-dimensional space can be used to visualize the key role of the dynamic coupling mechanism between meteorological factors and electricity price fluctuations in load forecasting. Experimental results show that load demand points in high temperature and high humidity areas (for example, temperatures exceeding 30°C and humidity greater than 70%) show a significant clustered upward trend, and the steep rise in the surface of this area highly coincides with the warm tones of the coupling characteristics, proving that the interaction between temperature and humidity causes a surge in load through the electricity price transmission mechanism. Especially in extreme meteorological areas, the coupling characteristics generated by the embodiment of the present invention accurately capture the nonlinear jump points of the load, indicating that the embodiment of the present invention scientifically quantifies the coupling strength by mixing partial derivatives and gradient norms, thereby solving the inherent defects of the existing technology in independently processing meteorological electricity price information, and can accurately capture the intensity and pattern of the impact of changes in meteorological conditions on electricity prices.

[0037] After completing the above data processing, a deep feature extraction network structure is constructed to generate a park load forecasting model. The data obtained after the above processing is sampled to perform a model sequence on the park load forecasting model for subsequent load demand forecasting of the target park. The specific model construction and training process is as follows: The model adopts a hierarchical encoder architecture, with an input layer, a bidirectional temporal attention block, a feature distillation layer, a decoding layer, and an output layer. During the training process, the input layer receives the concatenation of historical multi-load features and historical coupling features that have been processed by spatiotemporal alignment. The stacked bidirectional temporal attention blocks are the core of the model and the backbone of feature extraction. The output of each attention block is connected to the feature distillation layer. The first layer of the feature distillation layer uses a dilation rate of 2 to capture long-term patterns. The second layer uses maximum pooling with a stride of 2 to compress redundant features. The last layer dynamically weights key feature dimensions through a channel attention mechanism. The decoding layer (also called the decoding end) uses a gated deconvolution module to gradually upsample, and finally connects to a time-distributed fully connected network through the output layer to generate a prediction sequence.

[0038] In an embodiment of the present invention, considering that load demand forecasting needs to capture both the rapid transient response of a single flexible load entity and the long-term trend change of the entire park load, it is difficult for the traditional convolutional neural network and recurrent neural network combination model to effectively and uniformly model such long-term and short-term temporal dependencies. In addition, in temporal modeling, if handled improperly, it is easy to use information of future time steps to predict the current moment, resulting in future information leakage, causing the model to fail in actual prediction; based on this, an embodiment of the present invention adopts a bidirectional temporal attention block; for each bidirectional temporal attention block, first, a one-dimensional causal convolution is used to process the input feature matrix. The causal convolution ensures that the convolution output only depends on the input of the current and past moments through a specific left-filling method, strictly guarantees temporal causality, effectively prevents future information leakage, and is mainly used to extract forward temporal dependency features; at the same time, a one-dimensional time-reverse convolution is used to process the same input feature matrix. The input feature matrix is ​​first transformed along time. The axis is reversed, convolution is performed, and then the original time series is reversed to capture the backward historical pattern characteristics contained in the sequence. The output of the causal convolution is used as the query vector matrix, and the output of the time-reversed convolution is used as the key vector matrix and the value vector matrix respectively. Then, the query vector matrix and the key vector matrix are matrix-producted to calculate the attention score. After adjusting the scaling factor, the row is normalized using the Softmax function to obtain the attention weight, which is then applied to the value vector matrix to generate the output of the attention mechanism (i.e., the attention mechanism output matrix). After that, the attention output (i.e., the attention mechanism output matrix) and the output of the causal convolution (i.e., the query vector matrix) are residually connected. The forward dependency modeling capability of the causal convolution and the backward pattern extraction capability of the time-reversed convolution are jointly utilized. The bidirectional information is then fused through the attention mechanism, and the original features are retained through the residual connection to achieve unified modeling of complex long-term and short-term temporal dependencies in the load information. The specific representation is as follows: (5); (6); (7); in, Represents the query vector matrix, with dimension , generated by causal convolution to capture forward temporal dependencies; represents the parameter space; represents the key vector matrix, is the matrix identifier, and its dimensions are , generated by time-reversed convolution, and used to represent backward time series features; represents a matrix of value vectors, is the matrix identifier, and its dimensions are , generated synchronously with the key vector; represents the implicit dimension, i.e. The characteristic dimensions of ; Represents a one-dimensional causal convolution operation, In the parameter Indicates that the convolution kernel width is 5 time steps, and the left padding is used to ensure that the output time series length remains unchanged to prevent future information leakage; represents the input feature matrix, ; Represents a one-dimensional time-reversed convolution operation, In the parameter Indicates that the convolution kernel width is 5, the input sequence is first reversed along the time axis, and then reversed back to the original time sequence after convolution; Indicates the size of the convolution kernel and controls the receptive field of feature extraction; Represents the attention mechanism output matrix, with dimension ; represents the row normalized exponential function, which converts the attention score into a probability distribution; represents the transpose of the product of the query vector matrix and the key vector matrix; Represents a scaling factor used to stabilize the gradient.

[0039] For the model constructed above, the modeling advantage of the bidirectional temporal attention block for complex load patterns is verified by comparing multi-scale curves. The prediction results of the embodiment of the present invention can closely match the long-term trend of the actual load. The embodiment of the present invention can accurately capture the 300kW peak fluctuation lasting 25 minutes, while the comparison model (such as Transformer and LSTM) has response delay and amplitude attenuation. In addition, the advantage is further quantified by comparing the absolute error curves, such as Figure 2As shown in the figure, during the device state switching period (such as the 12 / 36th hour), the error of the embodiment of the present invention is always less than 50kW, while the error of the comparison model soars to more than 120kW, proving that the combined mechanism of bidirectional causal convolution and time reverse convolution can effectively unify the transient response and long-term trend modeling capabilities.

[0040] Furthermore, considering that the load information of the park power grid usually shows obvious time period patterns, such as different patterns on weekdays and holidays, and different load curves in different seasons, if the model uses conventional random initialization weights, the model needs to learn these complex patterns from scratch, which will cause the training process to converge slowly; based on this, after completing the model construction, the embodiment of the present invention adopts an initialization method based on historical load pattern clustering to perform adaptive weight initialization, so that the model can start learning from a typical load pattern and accelerate convergence. The specific process is as follows: First, the number of cluster centers is preset, and the k-means clustering algorithm is used to analyze the historical input feature matrix set, and the historical input feature matrix is ​​divided into several categories. The cluster center of each category represents a typical load pattern, such as the corresponding seasonal patterns. (such as winter, summer, etc.) or time mode (such as peak time) or seasonal and time combination mode; then, calculate the variance of each cluster to characterize the stability of the load pattern of this type; then, according to the proportion of the variance of each cluster to the total variance of all clusters, obtain the weight coefficient of the class. Clusters with larger variances usually contain more diverse patterns and are given higher weights; perform principal component analysis dimensionality reduction on the high-dimensional load pattern represented by each cluster center, and reduce the dimension of the cluster center to the dimension consistent with the model input layer; use the weight coefficient of each cluster cluster to perform weighted summation on the features of each cluster center after dimensionality reduction to generate the final initialization weight matrix, which is used to initialize the parameters of the first layer constructed above, so that the initial state of the model contains typical historical load pattern knowledge, which is specifically expressed as follows: (8); (9); (10); in, For the Cluster centers, generated by the k-means clustering algorithm, represent the Class load mode, The load patterns corresponding to different seasons / time periods; is a set of historical input feature matrices, containing historical load information for multiple unit times; is the k-means clustering algorithm, Indicates that the number of cluster centers is 4; For the The weight coefficient of each cluster is determined by the ratio of the variance within the cluster to the total variance of the clusters; For the The variance of the cluster corresponding to each cluster center represents the stability of the load pattern. is the sum of the variances of the four clusters; To initialize the weight matrix, used to initialize the parameters of the first layer of the model; Represents the cluster center Perform principal component analysis to reduce dimensionality. The dimension after dimensionality reduction is consistent with the model input layer dimension.

[0041] Furthermore, considering that the load curve often undergoes step changes due to the start and stop of the flexible load body, the traditional ReLU family activation function is not differentiable at the input value of zero, which is not conducive to model optimization. At the same time, ReLU will set all negative input values ​​to zero, ignoring negative signals that may contain important information, which is not conducive to capturing load reduction or shutdown states. Based on this, the present invention adopts a smoothing rectifier unit activation function, combined with a dynamic adjustment threshold calculated by the mean of the absolute value of the input feature vector after splicing the historical multivariate load characteristics and historical coupling characteristics corresponding to all samples in the current training batch, to control the boundaries of different regions of the function. For the case where the input value is greater than the dynamic adjustment threshold, the function adopts a linear response to maintain normal transmission of positive signals. For the case where the input value is less than the negative dynamic adjustment threshold, the function adopts a linear response with a fixed negative leakage coefficient to retain key negative information. In addition, in the interval where the absolute value of the input value is less than or equal to the dynamic adjustment threshold, the function adopts a quadratic function form to ensure that the function is not only continuous at the zero point, but also its first-order derivative is continuous, achieving a smooth transition, so that the activation function is particularly suitable for step change points in the load information caused by the start and stop of the equipment, while maintaining computational efficiency, providing better gradient and negative information processing capabilities. The activation function of the model is expressed as follows: (11); in, Represents the neuron input value of the model, that is, the weighted sum of the output of the previous layer; is the activation function of the smoothing rectifier unit; It is a dynamic adjustment threshold value, which is calculated from the current batch data and controls the boundary of the linear interval. The calculation method is expressed as ; is the number of samples in the current training batch; For the The input feature matrix corresponding to the samples (can also be described as the input feature vector); The absolute value of the input feature matrix is ​​obtained by taking the absolute value of each element and then calculating the norm. Leakage coefficient in the negative area, such as taking a fixed value , control the degree of negative information retention; When , the activation function uses the quadratic function , make sure The first-order derivative is continuous and smoothly transitioned; When , the activation function is linear , maintains a linear response for positive values; When , the function is linear , retain negative value information.

[0042] Furthermore, considering that the use of single-objective loss functions such as mean absolute error or mean square error focuses on the accuracy of the predicted point value but ignores the smoothness of the predicted load curve and the consistency of the overall distribution characteristics of the predicted load with the actual load distribution; based on this, the present invention constructs a dual-penalty loss function containing three core terms in real time to simultaneously optimize prediction accuracy, curve smoothness, and distribution matching. The first term is the mean absolute error term, which directly measures the prediction accuracy by calculating the average of the absolute errors between the predicted values ​​in all historical park load forecast curves and the actual load values ​​at the corresponding time. The second term is the curvature constraint term. First, the second-order difference of the predicted load sequence corresponding to the historical park load forecast curve is calculated to characterize the curvature or smoothness change of the historical park load forecast curve, that is, the smoothness of the load curve. Then, the Frobenius norm of this second-order difference matrix is ​​calculated as the curvature constraint term (i.e., penalty term) to suppress the occurrence of drastic and unrealistic fluctuations in the predicted load curve. The third term is the distribution matching term, which calculates the Kullback-Leibler (KL) divergence distribution between the probability distribution of the predicted load sequence corresponding to the historical campus load forecast curve and the probability distribution of the actual load sequence corresponding to the actual campus load curve at the corresponding time. The degree of difference between the two distributions is measured as the KL divergence distribution matching fluctuation (i.e., distribution matching) term, which is used to ensure that the overall statistical characteristics of the predicted load curve are consistent with the actual load curve. The final loss value can be obtained by summing the mean absolute error term, curvature constraint term, and distribution matching term calculated above. The overall dual-penalty loss function of the model is specifically expressed as follows: (12); in, is the overall loss value, which needs to be minimized through back propagation; is the forecast load vector at all times; for The predicted load value at the time; ; for The actual load value at the moment; for The absolute error between the predicted value and the true value at a given moment; is the mean absolute error term, Indicates the time length of the forecast / real load series; for The second-order difference of is calculated as ; is the Frobenius norm; is the curvature constraint penalty coefficient, such as a fixed value ; is the KL divergence distribution; is the probability distribution of the predicted load series; is the probability distribution of the real load sequence; The term represents the distribution matching term, which measures the difference between the predicted distribution and the true distribution; is the distribution matching penalty coefficient, such as taking a fixed value .

[0043] The prediction accuracy of the double penalty loss function, mean square error loss function MSE, and mean absolute error loss function MAE is verified through multi-dimensional analysis. In the time zone of the equipment cluster start-stop event, the conventional mean square error loss prediction loses 300kW mutation details due to over-smoothing, and the mean absolute error loss produces step-like artifacts. However, the prediction of the double penalty loss function in the embodiment of the present invention accurately reproduces the change slope of 120kW / minute. Figure 3 The probability density diagram shown reveals that the predicted load distribution of the embodiment of the present invention has a degree of overlap with the actual load distribution of more than 90%, which is significantly higher than the matching degree of the mean square error loss function MSE and the mean absolute error loss function MAE, proving that the KL divergence constraint effectively ensures the consistency of statistical characteristics; at the same time, by comparing the predicted load curves under the three loss functions, it can be seen that the load prediction curve of the embodiment of the present invention maintains the sharpness of the event response while the curvature fluctuation amplitude (0-0.03) is closest to the actual load, while the comparison method still has high-frequency oscillations in the stable period (curvature>0.04), confirming that the curvature constraint term successfully suppresses non-physical fluctuations.

[0044] Furthermore, in the later stage of model training, it is necessary to fine-tune the model parameters to achieve optimal performance. However, the park load information usually has non-stationary characteristics, such as pattern drift. The conventional fixed attenuation strategy cannot be dynamically adjusted according to the current actual learning state of the model, and it is difficult to adapt to this data characteristic, which may lead to poor convergence or oscillation. Based on this, the embodiment of the present invention adopts a variance-sensitive attenuator to dynamically adjust the attenuation speed of the learning rate. The adjustment is based on the fluctuation of the loss value corresponding to the verification data set (for example, 70% of the original data set is divided into a training data set for model training, and 30% is divided into a verification data set for model verification) in the recent consecutive training rounds. The verification loss variance is obtained by calculating the loss value of the verification data set in the most recent rounds, reflecting the stability of the model performance on the verification data set. In each round At the beginning of training, the learning rate of this round is calculated by the natural exponential function based on the ratio of the current training round to the preset total number of training rounds and the current verification loss variance value. Specifically, when the loss value shows a high variance, it means that the model training may be unstable or in a sensitive area. At this time, the decay rate of the learning rate will be accelerated to help stabilize the training process. Conversely, when the loss value shows a low variance, it means that the model training is in good condition. At this time, the strategy will slow down the decay rate of the learning rate, allowing the model to make more refined parameter adjustments at the current learning rate to pursue a better solution. In addition, the initial learning rate is used as the starting point, and the variance normalization factor is used to balance the impact of the magnitude of the loss value variance on the decay calculation, so as to achieve adaptive matching of the learning rate decay with the actual training state and training progress of the model. The specific expression of the model's learning rate adjustment is as follows: (13); in, For the The learning rate during the training round; is the initial learning rate, such as a fixed value ; is the natural exponential function; is the variance of the loss value, which is calculated by the loss value of the last five consecutive training rounds; is the loss value; is the variance normalization factor, such as taking a fixed value , used to balance the magnitude of variance; The current training round, starting from 1; is the preset total number of training rounds; is the training progress coefficient, ranging from [0,1], representing the current training completion ratio.

[0045] In the embodiment of the present invention, mini-batch gradient descent is used in the training process, and a double convergence determination is performed using a validation dataset after each round of training; Specifically, the system first checks whether the loss value on the validation dataset has decreased by less than the decay threshold set by the dynamic learning rate scheduler for five consecutive rounds. Secondly, it checks whether the KL divergence distribution match between the load forecast sequence and the true composite sequence (i.e., the KL divergence distribution match fluctuation) meets a preset fluctuation threshold (i.e., a stability bound, such as 1%). When these convergence conditions are met simultaneously, the early stopping mechanism is triggered to save the optimal weights and output the trained model. If the convergence conditions are not met even after the training round limit is reached, the system automatically activates backup mode to load the historical best checkpoint (i.e., the optimal weights). The feature importance analysis module is then triggered to identify outdated features for targeted retraining. For example, methods such as SHAP values ​​and gradient analysis are used to identify features that have minimal impact on model accuracy or even do not affect model performance after deletion. These features are also considered outdated, indicating that they have no positive impact on training and may slow training by increasing computational complexity. After identifying the outdated features, the model is retrained after filtering them out from the historical input feature matrix to optimize training results.

[0046] In an optional embodiment, the preprocessing of the operating data and the meteorological information to obtain a multivariate load characteristic includes: Calculating the contribution of each flexible load subject according to the absolute deviation between the first historical average power consumption of each flexible load subject and the second average power consumption of the target park and a preset weight distribution steepness adjustment parameter; Performing time alignment on the load information in the operation data of each of the flexible load bodies using a dynamic time warping function to obtain the load characteristics of each of the flexible load bodies; Calculating the load characteristics of the target park according to the load characteristics and contribution of each of the flexible load entities; Performing spatial interpolation diffusion on the meteorological information using a preset convolution kernel to obtain meteorological characteristics of the meteorological information after spatial interpolation within the target park; Performing time alignment and interpolation processing on the electricity price information and the load time series characteristics to obtain the electricity price characteristics; The load characteristics, the meteorological characteristics and the electricity price characteristics of the target park are combined to obtain the multivariate load characteristics.

[0047] In an optional embodiment, dynamically coupling the correlation between meteorological conditions and electricity price fluctuations based on the multi-load characteristics to obtain coupling characteristics includes: Calculating a mixed second-order partial derivative of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating the gradient Euclidean norm of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating a meteorological coupling weight based on a nonlinear activation function according to the mixed second-order partial derivative, the gradient Euclidean norm, and a preset sensitivity coefficient; Calculating the product of the temperature element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a temperature-electricity price product feature; Calculating the product of the humidity element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a humidity-electricity price product feature; The coupling feature is generated according to the meteorological coupling weight, the temperature electricity price product feature, and the humidity electricity price product feature.

[0048] For example, after completing the model training, the operational data (including load information and electricity prices) of each flexible load entity within the target park within a preset unit time period, as well as meteorological information monitored by meteorological stations surrounding the target park, are collected in real time using the aforementioned preprocessing method to construct a real-time multivariate load signature. The aforementioned meteorological-electricity price coupling method is then used to generate a real-time coupled signature. The resulting real-time multivariate load signature and coupled signature are then concatenated and input into the trained park load forecasting model. A bidirectional temporal attention block captures long-term and short-term dependencies. Ultimately, the output layer generates a park load forecast curve for a predetermined sliding time window in the future, indicating the load resource demand for the park within the predetermined sliding time window. Finally, based on this park load forecast curve, load resource adjustments are made to the flexible load entities within the target park (e.g., those marked as adjustable) to balance the load resources within the park, forming a closed-loop "forecast-optimization" control chain. For example, the model performs a rolling forecast update every 15 minutes, and the model hot update mechanism is automatically triggered by the model stability indicator recorded by the dynamic learning rate scheduler.

[0049] For example, by comparing the error fluctuations of the method of the embodiment of the present invention and other methods at different time periods within a day, we pay special attention to the prediction stability during the peak power consumption period. The comparison results are as follows: Figure 4 As shown, Figure 4The horizontal axis represents a 24-hour period, and the vertical axis shows the mean absolute error values ​​under different methods, characterizing the prediction performance of the autoregressive moving average model ARIMA, the long short-term memory network LSTM, the prophet model Prophet, the convolutional neural network-gated recurrent unit combination CNN-GRU, and the implementation of the present invention. The experimental results show that the errors of all methods increase significantly during the peak hours of electricity consumption in the morning and evening (such as 7-9 o'clock and 17-19 o'clock), but the embodiment of the present invention always maintains the lowest error level with the smallest fluctuation. Especially in the afternoon when the air-conditioning load surges, the embodiment of the present invention accurately captures the dynamic interaction effect of temperature, humidity and electricity price through the meteorological-electricity price coupling feature generation mechanism, significantly reducing the peak load prediction deviation. The error curve of the embodiment of the present invention is moderately smooth and has reasonable fluctuations, reflecting the complexity of load changes in real scenarios. Its outstanding performance in key periods verifies the advantages of the spatiotemporal dynamic alignment layer in fusing multi-source heterogeneous data.

[0050] By further comparing the autoregressive moving average model ARIMA, the long short-term memory network LSTM, the prophet model Prophet, the convolutional neural network-gated recurrent unit combination CNN-GRU and the scatter distribution diagram implemented by the present invention, the correlation between the prediction results and the actual load value can be tested, such as Figure 5-8 As shown in the figure, the horizontal axis is the actual load value, the vertical axis is the predicted load value, and the reference diagonal line represents the ideal prediction. It can be seen that the scatter points implemented by the present invention are closely distributed on both sides of the reference line, forming a narrow and concentrated band distribution, while the scatter points of other methods show a diffuse trend, proving that the prediction results implemented by the present invention have the smallest systematic deviation. This shows that the curvature constraint term suppresses unrealistic load fluctuations, and the distribution matching term ensures that the predicted sequence is consistent with the statistical characteristics of the actual load. Especially in high-load areas, the scatter points implemented by the present invention still maintain a good linear relationship. The park load forecasting model of the present invention has a stronger ability to predict peak load.

[0051] Compared with the existing technology, the embodiment of the present invention integrates data with different sampling rates and different spatial granularities through dynamic spatiotemporal alignment operations, ensuring the temporal correlation of the load sequences of various devices in the park. At the same time, it overcomes the limitations of conventional interpolation and downsampling methods, avoids information loss, and can process heterogeneous data from different devices and sensors. It improves the system's adaptability to changes in the load demand of the park power grid, especially in highly dynamic and complex environments, and can maintain high prediction accuracy and stability; at the same time, by calculating the mixed second-order partial derivatives of temperature and humidity and the sensitivity of electricity prices to meteorological gradients, it dynamically quantifies the nonlinear coupling effect between meteorology and electricity prices, captures the impact of changes in temperature, humidity and electricity prices on load fluctuations, and solves the problem of conventional methods ignoring this interaction effect, significantly reducing the prediction error during peak hours, especially in scenarios where equipment load surges, effectively predicting load fluctuations, and further improving the accuracy of park load forecasting.

[0052] The embodiment of the present invention constructs a bidirectional temporal attention block and jointly extracts temporal dependencies through causal convolution and time-reversal convolution, which can uniformly model the transient response and long-term trend in load information, avoid future information leakage, and solve the problem of insufficient modeling of short-term and long-term temporal dependencies in traditional models, thereby improving prediction accuracy and stability, and having significant advantages in dealing with complex patterns of campus loads.

[0053] The embodiment of the present invention performs adaptive weight initialization of the model based on the historical load pattern obtained by clustering, avoids learning complex load patterns from scratch, accelerates model convergence, and introduces a variance-sensitive attenuator strategy to dynamically adjust the learning rate, thereby optimizing the training process. In particular, it is more adaptable to non-stationary load information and adapts to the time-varying characteristics of park load information.

[0054] See also Figure 9 , Figure 9 The embodiment of the present invention provides a structural block diagram of a campus power grid flexible load resource optimization system based on deep learning, and the campus power grid flexible load resource optimization system based on deep learning includes: The data acquisition module 11 is used to collect the operating data of each flexible load entity in the target park and the meteorological information monitored by the meteorological station around the target park; wherein the operating data includes the load information and electricity price information of the corresponding flexible load entity; a first data preprocessing module 12 for preprocessing the operating data and the meteorological information to obtain a multi-element load characteristic; wherein the multi-element load characteristic includes the load characteristic, meteorological characteristic, and electricity price characteristic of each of the flexible load entities after time and space alignment; A first feature coupling module 13 is configured to dynamically couple the correlation between meteorological conditions and electricity price fluctuations based on the multi-load characteristics to obtain coupling characteristics; A load forecasting module 14 is configured to perform load forecasting using a pre-trained park load forecasting model based on the multi-element load characteristics and the coupling characteristics, and obtain a park load forecast curve within a future set time window; The load resource adjustment module 15 is configured to adjust the load resources of the flexible load entities in the target park according to the park load forecast curve.

[0055] In an optional embodiment, the first data preprocessing module 12 includes: a contribution calculation unit, configured to calculate the contribution of each of the flexible load entities based on an absolute deviation between the first historical average power consumption of each of the flexible load entities and the second average power consumption of the target park and a preset weight distribution steepness adjustment parameter; a first load characteristic generating unit, configured to perform time alignment on load information in the operation data of each of the flexible load bodies using a dynamic time warping function to obtain a load characteristic of each of the flexible load bodies; A second load characteristic generating unit is configured to calculate the load characteristic of the target park according to the load characteristic and contribution of each of the flexible load bodies; A meteorological feature generating unit, configured to perform spatial interpolation diffusion on the meteorological information using a preset convolution kernel to obtain meteorological features of the meteorological information after spatial interpolation within the target park; an electricity price feature generating unit, configured to perform time alignment and interpolation processing on the electricity price information and the load time series feature to obtain an electricity price feature; The multi-element load characteristic generating unit is used to combine the load characteristic of the target park, the meteorological characteristic and the electricity price characteristic to obtain the multi-element load characteristic.

[0056] In an optional embodiment, the first feature coupling module 13 includes: a mixed second-order partial derivative calculation unit, configured to calculate a mixed second-order partial derivative of the electricity price characteristic with respect to the meteorological characteristic based on the multivariate load characteristic; a gradient Euclidean norm calculation unit, configured to calculate a gradient Euclidean norm of the electricity price characteristic with respect to the meteorological characteristic based on the multivariate load characteristic; a meteorological coupling weight calculation unit, configured to calculate a meteorological coupling weight based on a nonlinear activation function according to the mixed second-order partial derivative, the gradient Euclidean norm, and a preset sensitivity coefficient; a temperature-electricity-price product feature generating unit, configured to calculate the product of the temperature element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature, to obtain the temperature-electricity-price product feature; a humidity-electricity-price product feature generating unit, configured to calculate the product of the humidity element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a humidity-electricity-price product feature; A coupling feature generating unit is configured to generate the coupling feature according to the meteorological coupling weight, the temperature-electricity-price product feature, and the humidity-electricity-price product feature.

[0057] In an optional embodiment, the system further includes: A historical data collection module is used to collect historical operating data of each flexible load entity in the target park and historical meteorological information monitored by meteorological stations around the target park; wherein the historical operating data includes historical load information and historical electricity price information of the corresponding flexible load entity; a second preprocessing module, configured to preprocess the historical operating data and the historical meteorological information to obtain historical multi-element load characteristics; wherein the historical multi-element load characteristics include historical load characteristics, historical meteorological characteristics, and historical electricity price characteristics; A second feature coupling module is used to dynamically couple the correlation between meteorological conditions and electricity price fluctuations based on the historical multi-load characteristics to obtain historical coupling characteristics; The model training module is used to train the park load forecasting model according to the historical multi-load characteristics and the historical coupling characteristics until the preset convergence conditions are met, thereby obtaining a trained park load forecasting model.

[0058] In an optional embodiment, the model training module includes: a feature concatenation unit, configured to concatenate the historical multivariate load feature and the historical coupling feature to obtain an input feature matrix; A training unit, configured to input the input feature matrix into a park load forecasting model constructed based on bidirectional temporal attention to perform model training, and obtain a historical park forecast load curve output in this round; A loss calculation unit, configured to calculate the loss value and KL divergence distribution matching fluctuation between the historical park forecast load curve and the corresponding park actual load curve using a preset double-penalty loss function; A convergence judgment unit is used to judge whether the convergence condition is met based on the loss value and the KL divergence distribution matching fluctuation; wherein, the convergence condition includes: the decrease in the loss value for multiple consecutive rounds is less than a preset attenuation threshold, and the KL divergence distribution matching fluctuation for multiple consecutive rounds is less than a preset fluctuation threshold; if so, stop training the park load prediction model and output the trained park load prediction model; if not, continue training the park load prediction model.

[0059] In an optional embodiment, the park load forecasting model includes an input layer arranged according to a hierarchical encoder architecture, a bidirectional temporal attention block, a feature distillation layer, a decoding layer constructed based on a gated deconvolution module, and an output layer; The forward temporal dependency features of the input feature matrix received by the input layer are captured by the causal convolution layer in the bidirectional temporal attention block to obtain a query vector matrix; Capturing backward historical pattern features of the input feature matrix received by the input layer through the time-reversed convolution layer in the bidirectional temporal attention block to obtain a key vector matrix and a value vector matrix; In the attention mechanism layer of the bidirectional temporal attention block, attention weights are calculated according to the query vector matrix and the key vector matrix, and an attention mechanism output matrix is ​​generated according to the attention weights, the value vector matrix, and the query vector matrix; The attention mechanism output matrix is ​​processed in sequence by the feature distillation layer, the decoding layer, and the output layer to generate a corresponding historical park predicted load curve.

[0060] In an optional embodiment, the system further includes: A cluster analysis module, configured to perform a k-means cluster analysis on the input feature matrix to divide the input feature matrix into a plurality of clusters, wherein the cluster center of each cluster corresponds to a load mode; The clustering weight coefficient calculation module is used to calculate the variance of each cluster and obtain the weight coefficient of the corresponding cluster according to the proportion of the variance of each cluster to the total variance of all clusters. A dimensionality reduction module, configured to perform a principal component analysis dimensionality reduction operation on each of the cluster centers so that the dimension of each cluster center meets the input requirements of the park load forecasting model; A weight matrix calculation module is used to perform weighted summation on the cluster centers after dimensionality reduction according to the weight coefficients of the clusters to obtain an initialized weight matrix; The model parameter initialization module is used to initialize the model parameters of the park load forecasting model using the initialization weight matrix.

[0061] In an optional embodiment, the dual-penalty loss function includes a mean absolute error term, a curvature constraint term, and a distribution matching term; The loss calculation unit includes: a mean absolute error calculation subunit, configured to calculate the average of the absolute errors between the predicted load in the historical park predicted load curve and the actual load in the corresponding park actual load curve, to obtain a mean absolute error term; a norm calculation subunit, configured to calculate a second-order difference of the predicted load in the historical park predicted load curve, and calculate a Frobenius norm of the second-order difference to obtain a curvature constraint term; A divergence distribution calculation subunit, configured to calculate a KL divergence distribution matching fluctuation between the probability distribution of the historical park predicted load curve and the probability distribution of the corresponding park actual load curve as a distribution matching item; The loss value calculation subunit is used to obtain the loss value between the historical park predicted load curve and the corresponding park actual load curve based on the mean absolute error term, the curvature constraint term and the distribution matching term.

[0062] In an optional embodiment, the system further includes: The decay rate calculation module is used to calculate the decay rate of the learning rate of the current training round through the natural exponential function at the beginning of each training round based on the ratio of the current training round to the preset total number of training rounds and the variance of the loss value of the current training round; A learning rate adjustment module is used to dynamically adjust the learning rate of the park load forecasting model using the attenuation rate.

[0063] It should be noted that the working process of each module in the deep learning-based campus power grid flexible load resource optimization system described in the embodiment of the present invention can refer to the working process of the deep learning-based campus power grid flexible load resource optimization method described in the above embodiment, and the technical effect achieved is also the same as the deep learning-based campus power grid flexible load resource optimization method described in the above embodiment, which will not be repeated here.

[0064] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, various improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for optimizing flexible load resources in a park power grid based on deep learning, characterized in that: include: Collecting operating data of each flexible load entity in the target park and meteorological information monitored by meteorological stations around the target park; wherein the operating data includes load information and electricity price information of the corresponding flexible load entity; Preprocessing the operating data and the meteorological information to obtain a multi-element load characteristic; wherein the multi-element load characteristic includes a load characteristic, a meteorological characteristic, and an electricity price characteristic; According to the multi-load characteristics, the correlation between the weather and the electricity price fluctuation is dynamically coupled to obtain the coupling characteristics; Based on the multi-element load characteristics and the coupling characteristics, load forecasting is performed using a pre-trained park load forecasting model to obtain a park load forecast curve within a future set time window; According to the park load forecast curve, load resources of the flexible load entities in the target park are adjusted.

2. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 1, characterized in that: The preprocessing of the operating data and the meteorological information to obtain multivariate load characteristics includes: Calculating the contribution of each flexible load subject according to the absolute deviation between the first historical average power consumption of each flexible load subject and the second average power consumption of the target park and a preset weight distribution steepness adjustment parameter; Performing time alignment on the load information in the operation data of each of the flexible load bodies using a dynamic time warping function to obtain the load characteristics of each of the flexible load bodies; Calculating the load characteristics of the target park according to the load characteristics and contribution of each of the flexible load entities; Performing spatial interpolation diffusion on the meteorological information using a preset convolution kernel to obtain meteorological characteristics of the meteorological information after spatial interpolation within the target park; Performing time alignment and interpolation processing on the electricity price information and the load time series characteristics to obtain the electricity price characteristics; The load characteristics, the meteorological characteristics and the electricity price characteristics of the target park are combined to obtain the multivariate load characteristics.

3. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 1, characterized in that: The dynamic coupling of the correlation between meteorological conditions and electricity price fluctuations based on the multi-load characteristics to obtain coupling characteristics includes: Calculating a mixed second-order partial derivative of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating the gradient Euclidean norm of the electricity price characteristic with respect to the meteorological characteristic according to the multivariate load characteristic; Calculating a meteorological coupling weight based on a nonlinear activation function according to the mixed second-order partial derivative, the gradient Euclidean norm, and a preset sensitivity coefficient; Calculating the product of the temperature element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a temperature-electricity price product feature; Calculating the product of the humidity element at each moment in the meteorological feature and the electricity price element at the corresponding moment in the electricity price feature to obtain a humidity-electricity price product feature; The coupling feature is generated according to the meteorological coupling weight, the temperature electricity price product feature, and the humidity electricity price product feature.

4. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 1, characterized in that: The training process of the park load forecasting model includes: Collecting historical operating data of each flexible load entity in the target park and historical meteorological information monitored by meteorological stations around the target park; wherein the historical operating data includes historical load information and historical electricity price information of the corresponding flexible load entity; Preprocessing the historical operating data and the historical meteorological information to obtain historical multi-element load characteristics; wherein the historical multi-element load characteristics include historical load characteristics, historical meteorological characteristics, and historical electricity price characteristics; According to the historical multi-load characteristics, the correlation between meteorological conditions and electricity price fluctuations is dynamically coupled to obtain historical coupling characteristics; According to the historical multi-load characteristics and the historical coupling characteristics, the park load forecasting model is trained until a preset convergence condition is met, thereby obtaining a trained park load forecasting model.

5. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 4, characterized in that: The method of training the park load forecasting model according to the historical multi-load characteristics and the historical coupling characteristics until a preset convergence condition is met to obtain a trained park load forecasting model includes: splicing the historical multivariate load characteristics and the historical coupling characteristics to obtain an input characteristic matrix; Inputting the input feature matrix into the park load forecasting model built based on bidirectional temporal attention for model training, and obtaining the historical park forecast load curve output in this round; Calculate the loss value and KL divergence distribution matching fluctuation between the historical park forecast load curve and the corresponding park real load curve through a preset double penalty loss function; Determining whether the convergence condition is satisfied based on the loss value and the KL divergence distribution matching fluctuation; wherein the convergence condition includes: the decrease in the loss value for multiple consecutive rounds is less than a preset attenuation threshold, and the KL divergence distribution matching fluctuation for multiple consecutive rounds is less than a preset fluctuation threshold; If so, stop training the park load forecasting model and output the trained park load forecasting model; If not, continue training the park load forecasting model.

6. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 5, characterized in that: The park load forecasting model includes an input layer, a bidirectional temporal attention block, a feature distillation layer, a decoding layer based on a gated deconvolution module, and an output layer arranged according to a hierarchical encoder architecture; The forward temporal dependency features of the input feature matrix received by the input layer are captured by the causal convolution layer in the bidirectional temporal attention block to obtain a query vector matrix; Capturing backward historical pattern features of the input feature matrix received by the input layer through the time-reversed convolution layer in the bidirectional temporal attention block to obtain a key vector matrix and a value vector matrix; In the attention mechanism layer of the bidirectional temporal attention block, attention weights are calculated according to the query vector matrix and the key vector matrix, and an attention mechanism output matrix is ​​generated according to the attention weights, the value vector matrix, and the query vector matrix; The attention mechanism output matrix is ​​processed in sequence by the feature distillation layer, the decoding layer, and the output layer to generate a corresponding historical park predicted load curve.

7. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 5, characterized in that: The method further comprises: Performing a k-means cluster analysis on the input feature matrix to divide the input feature matrix into a plurality of clusters, wherein the cluster center of each cluster corresponds to a load pattern; Calculate the variance of each cluster, and obtain the weight coefficient of the corresponding cluster according to the proportion of the variance of each cluster to the total variance of all clusters. Performing a principal component analysis dimensionality reduction operation on each of the cluster centers so that the dimensions of each of the cluster centers meet the input requirements of the park load forecasting model; According to the weight coefficients of the clusters, weighted summation is performed on the cluster centers after dimensionality reduction to obtain an initialization weight matrix; The initialization weight matrix is ​​used to initialize the model parameters of the park load forecasting model.

8. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 5, characterized in that: The dual penalty loss function includes a mean absolute error term, a curvature constraint term, and a distribution matching term; The calculation of the loss value and KL divergence distribution matching fluctuation between the historical park predicted load curve and the corresponding park actual load curve using a preset double penalty loss function includes: Calculating the average of the absolute errors between the predicted load in the historical park predicted load curve and the actual load in the corresponding park actual load curve to obtain a mean absolute error term; Calculating the second-order difference of the predicted load in the historical park predicted load curve, and calculating the Frobenius norm of the second-order difference to obtain a curvature constraint term; Calculating the KL divergence distribution matching fluctuation between the probability distribution of the historical park predicted load curve and the probability distribution of the corresponding park actual load curve as a distribution matching item; According to the mean absolute error term, the curvature constraint term and the distribution matching term, a loss value between the historical park predicted load curve and the corresponding park actual load curve is obtained.

9. The method for optimizing flexible load resources of a campus power grid based on deep learning according to claim 8, characterized in that: The method further comprises: At the beginning of each round of training, the decay rate of the learning rate of the current training round is calculated by the natural exponential function based on the ratio of the current training round to the preset total number of training rounds and the variance of the loss value of the current training round; The decay rate is adopted to dynamically adjust the learning rate of the park load forecasting model.

10. A campus power grid flexible load resource optimization system based on deep learning, characterized in that: include: A data acquisition module, configured to collect operating data of each flexible load entity within a target park and meteorological information monitored by meteorological stations surrounding the target park; wherein the operating data includes load information and electricity price information of the corresponding flexible load entity; a data preprocessing module, configured to preprocess the operating data and the meteorological information to obtain a multivariate load characteristic; wherein the multivariate load characteristic includes the load characteristic, meteorological characteristic, and electricity price characteristic of each of the flexible load entities after time and space alignment; A feature coupling module is used to dynamically couple the correlation between meteorological conditions and electricity price fluctuations according to the multi-load characteristics to obtain coupling features; A load forecasting module, configured to perform load forecasting using a pre-trained park load forecasting model based on the multi-element load characteristics and the coupling characteristics, and obtain a park load forecast curve within a future set time window; The load resource adjustment module is used to adjust the load resources of the flexible load body in the target park according to the park load forecast curve.

Images