Power grid load prediction method and system based on deep learning

Abstract

The application discloses a power grid load prediction method and system based on deep learning, which comprises the following steps: collecting multi-dimensional load correlation data through an intelligent power grid data fusion calculation platform, cooperatively capturing space-time correlation characteristics through a space-time residual gated recurrent model, enhancing network and strengthening key feature representation through a gated recurrent unit, then inputting a multi-modal load prediction deep model to build feature mapping relationship, generating preliminary prediction results through feature weight distribution, abnormal value elimination and smoothing optimization, and outputting final data after platform verification. The space-time residual gated recurrent model, the gated recurrent unit enhancement network and the multi-modal load prediction deep model all have exclusive feature processing mechanisms, and the processing accuracy is guaranteed through feature classification, gate initialization, mode division, fusion calculation and other subdivided processes. The application effectively improves the comprehensiveness and reliability of load prediction, and is suitable for accurate scheduling and optimized operation of intelligent power grids.

Description

Technical Field

[0001] This invention relates to the field of power grid load forecasting technology, and in particular to a power grid load forecasting method and system based on deep learning. Background Technology

[0002] With the deepening of smart grid construction, the scale of power grid operation continues to expand, and the spatiotemporal distribution characteristics of electricity load are becoming increasingly complex. Meanwhile, factors such as the integration of new energy sources and the diversification of electricity consumption patterns further exacerbate the uncertainty of load fluctuations. Accurate power grid load forecasting, as a core support for power dispatching, resource optimization, and the safe and stable operation of the power grid, is becoming increasingly important. Currently, smart grids have established a multi-dimensional data acquisition system, accumulating massive amounts of historical load data, equipment operating status data, regional electricity consumption characteristic data, and environmental influencing factor data, providing a rich data foundation for the application of deep learning technology. Deep learning-based power grid load forecasting methods and systems, relying on spatiotemporal residual gated cyclic models, gated cyclic unit enhanced networks, and multimodal load forecasting deep models, combined with the efficient data processing capabilities of smart grid data fusion computing platforms, have become a key technical path for solving complex load forecasting problems.

[0003] Existing technologies in the field of power grid load forecasting have two significant drawbacks: First, traditional forecasting methods fail to fully exploit the spatiotemporal correlation characteristics in load data, considering only time series dependence or spatial distribution patterns and ignoring their synergistic effect. This results in insufficient dynamic capture capability of load changes and difficulty in adapting to the complex spatiotemporal evolution characteristics of power grid loads. Second, existing models are not perfect in their fusion processing of multimodal data, failing to effectively address the heterogeneity of data from different sources and of different types. They lack targeted feature enhancement and fusion mechanisms, and their parameter optimization strategies are relatively simple, making it difficult to balance forecast accuracy and model stability. This makes the forecast results susceptible to interference from noisy data and unable to meet the high accuracy and high reliability requirements of smart grid load forecasting. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a power grid load forecasting method and system based on deep learning.

[0005] The technical solution adopted in this invention is a power grid load forecasting method based on deep learning, comprising the following steps: S1, collecting multi-dimensional load correlation data during power grid operation through a smart grid data fusion computing platform, the data including historical load values, power grid equipment operating status data, regional electricity consumption characteristic data, and environmental influencing factor data; S2, inputting the collected multi-dimensional data into a spatiotemporal residual gated recurrent model, and using the spatiotemporal feature extraction module in the model to collaboratively capture the time series dependencies and spatial distribution correlation features in the data; S3, using a gated recurrent unit enhancement network to perform deep enhancement processing on the feature vector output by the spatiotemporal residual gated recurrent model, and strengthening the characterization ability of the calibrated load influence features by dynamically adjusting the gate mechanism parameters; S4, inputting the enhanced feature vector into a multimodal load forecasting deep model, and constructing a feature mapping relationship by combining the heterogeneity features of the multimodal data; S5, generating preliminary load forecasting results through the output layer of the multimodal load forecasting deep model, and optimizing and adjusting the forecasting results based on the built-in feature weight allocation mechanism of the model; S6, performing data verification and outputting the optimized load forecasting results through a smart grid data fusion computing platform to form the final power grid load forecasting data.

[0006] Furthermore, the expression for the spatiotemporal residual gated cyclic model is: ,in, The feature output of the spatial location of s at time t. It is the sigmoid activation function. This is an element-wise multiplication operation. This is the model weight matrix. For bias terms, The input data is the spatial location of s at time t. For residual link terms, for Historical characteristics of spatial location at time s For time t Neighborhood characteristics of spatial location.

[0007] Furthermore, the expression for the gated recurrent unit enhancement network is:

[0008] ,

[0009] ,

[0010] ,

[0011] ,

[0012] in, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. The input feature vector at time t, for Hide your status at all times. As a load characteristic enhancement factor, For weight parameters, This is the bias parameter.

[0013] Furthermore, the expression for the multimodal load prediction depth model is: ,in, This is the load forecast value. For the number of multimodal data types, The weighting coefficients for the m-th modal data are... For the multilayer perceptron corresponding to the m-th modality, For convolutional feature extraction operations, Let m be the feature matrix of the m-th mode. For attention mechanism computational operations, This is the fusion feature matrix for all modalities.

[0014] Furthermore, the data fusion expression of the smart grid data fusion computing platform is as follows:

[0015]

[0016] Feature data, The number of dimensions for data collection. The fusion weights for the k-th dimension data are... Data is collected for the k-th dimension. The fusion bias for the k-th dimension data, For feature splicing operations, This is a normalization operation.

[0017] Furthermore, the parameter optimization expression for the power grid load forecast is as follows: ,in, This is the optimal set of model parameters. The parameters to be optimized in the model For the sample size, Let i be the actual load value of the i-th sample. Let i be the predicted load value for the i-th sample. The regularization coefficient is . This is an L2 regularization term.

[0018] Further, S3 includes the following sub-steps: S31, obtaining the feature vector output by the spatiotemporal residual gated recurrent model, dividing the vector into a time-related feature subset and a spatially related feature subset according to the feature dimension, and identifying the two subsets through a feature classifier; S32, inputting the identified feature subsets into the input layer of the gated recurrent unit enhancement network, and setting the initial gate parameters according to the variance and mean of the feature subsets through the dynamic gate initialization module; S33, starting the gated recurrent unit enhancement network based on the initial gate parameters, filtering and retaining historical feature information through the update gate, and filtering noise from the current input features through the reset gate; S34, performing a nonlinear transformation on the filtered features through the hidden layer of the gated recurrent unit enhancement network, strengthening the correlation between features of different dimensions in conjunction with the feature interaction module, and outputting the enhanced feature vector.

[0019] Further, S4 includes the following sub-steps: S41, modally partitioning the feature vectors output by the gated recurrent unit enhancement network, and assigning them to the corresponding modality processing branches according to the data source and feature type; S42, each modality processing branch maps the input features to a feature space of a unified dimension through a dedicated feature transformation layer, and adjusts the feature scale of different branches using a modality adaptation function; S43, inputting the features output by different modality processing branches into the fusion layer of the multimodal load prediction deep model, and calculating the correlation strength between different modality features through a cross-modal attention mechanism; S44, performing weighted fusion of different modality features based on the correlation strength to generate a multimodal fusion feature matrix, which serves as the input for subsequent calculations of the multimodal load prediction deep model.

[0020] Further, step S5 includes the following sub-steps: S51, inputting the multimodal fusion feature matrix into the fully connected layer of the multimodal load forecasting deep model, and converting the fusion features into a predicted value vector through matrix multiplication; S52, calling the model's built-in feature weight allocation mechanism, calculating feature weight values ​​according to the contribution of different features to load forecasting, and adjusting the predicted value vector by weight; S53, applying range constraints to the weighted predicted value vector through the model's output adjustment layer, and eliminating abnormal predicted values ​​that exceed the reasonable load range; S54, using a sliding window smoothing strategy to locally optimize the adjusted predicted values, eliminate prediction fluctuations, and generate preliminary load forecasting results.

[0021] This system, based on deep learning, is a power grid load forecasting system. It includes: a multi-dimensional data acquisition and transmission unit, used to acquire power grid load-related data through sensors and data acquisition terminals deployed in the smart grid, convert the data into a standardized format, and transmit it to the smart grid data fusion and computing unit; a smart grid data fusion and computing unit, connected to the multi-dimensional data acquisition and transmission unit, receiving the standardized data, performing fusion processing, generating feature data in a unified format, and simultaneously receiving and verifying the prediction results; and a spatiotemporal residual gated cyclic feature extraction unit, connected to the smart grid data fusion and computing unit, which performs spatiotemporal correlation feature extraction on the fused feature data. The system captures and outputs preliminary feature vectors. A gated cyclic unit, connected to the spatiotemporal residual gated cyclic feature extraction unit, receives the preliminary feature vectors and strengthens the calibration features through a dynamic gating mechanism, outputting enhanced feature vectors. A multimodal load forecasting calculation unit, connected to the gated cyclic unit, performs multimodal fusion and forecasting calculations on the enhanced feature vectors, generating optimized load forecasting results. A forecasting result output and feedback unit, connected to the multimodal load forecasting calculation unit and the smart grid data fusion calculation unit, receives the verified final forecasting results and outputs them to the dispatch center through the power grid dispatching interface. Simultaneously, it feeds back the forecasting error data to different model units for parameter adjustment.

[0022] Beneficial Effects: This invention proposes a deep learning-based power grid load forecasting method and system. It integrates multi-dimensional load-related data using a smart grid data fusion computing platform, and then collaboratively captures the temporal dependence and spatial distribution correlation features in the data through a spatiotemporal residual gated cyclic model, effectively solving the problem of insufficient mining of spatiotemporal collaborative features of loads in traditional methods. Subsequently, the gating cyclic unit enhances the network's dynamic adjustment of gating parameters, strengthening the representation capability of key load impact features. Combined with a multimodal load forecasting deep model, it specifically handles data heterogeneity, constructing accurate feature mapping relationships. Simultaneously, through mechanisms such as feature weight allocation, outlier removal, and smoothing optimization, it compensates for the shortcomings of existing technologies, such as incomplete multimodal data fusion and single parameter optimization. At the system level, six functional units are sequentially connected and efficiently linked, realizing a closed-loop operation of the entire process from data acquisition and transmission, fusion computing, feature extraction, enhancement processing, prediction computing, and result output feedback. This ensures both the efficiency of data processing and the comprehensiveness of feature extraction, while improving the accuracy and reliability of prediction results, providing strong technical support for power dispatching, resource optimization, and safe and stable operation of smart grids. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0024] Figure 2 This is a flowchart of method step S3 of the present invention;

[0025] Figure 3 This is a flowchart of method step S4 of the present invention;

[0026] Figure 4 This is a flowchart of step S5 of the method of the present invention;

[0027] Figure 5 This is a diagram showing the system module composition of the present invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the deep learning-based power grid load forecasting method includes the following steps:

[0030] S1. Collect multi-dimensional load-related data during the operation of the power grid through the smart grid data fusion computing platform. The data includes historical load values, power grid equipment operating status data, regional electricity consumption characteristic data, and environmental influencing factor data.

[0031] Specifically, step S1 involves the collection of multi-dimensional load-related data through a smart grid data fusion computing platform. This platform is equipped with 10 parallel data acquisition channels, supports a data transmission rate of 100MB per second, and can simultaneously connect to more than 2,000 power grid data acquisition terminals distributed in different areas. The collected multi-dimensional data includes four core categories: historical load data collected at 15-minute intervals, covering continuous operating data for the past 365 days, with a data accuracy of 0.01kW; power grid equipment operating status data, including 12 key parameters such as transformer load rate, line current, and switch status, collected once per minute; regional electricity consumption characteristic data, including 6 categories such as residential electricity consumption ratio, industrial electricity consumption ratio, and commercial electricity consumption ratio, updated at an hourly granularity; and environmental impact factor data, including 8 meteorological parameters such as temperature, humidity, wind speed, and precipitation, and 2 social factor data such as holiday type and major event arrangements. Meteorological parameters are collected at 10-minute intervals, and social factor data is updated in real time. All collected data is transmitted to the platform's data storage module via an encrypted transmission protocol. The storage module adopts a distributed architecture, supports petabyte-level data storage capacity, ensures data integrity and security, and provides comprehensive and accurate raw data support for subsequent model calculations.

[0032] S2 inputs the collected multi-dimensional data into the spatiotemporal residual gated cyclic model, and uses the spatiotemporal feature extraction module in the model to collaboratively capture the time series dependencies and spatial distribution correlation features in the data.

[0033] Specifically, step S2 organizes the multi-dimensional data collected and stored in S1 according to a preset data format specification and inputs it into a spatiotemporal residual gated recurrent model. This model includes an input layer, a spatiotemporal feature extraction module, a hidden layer, and an output layer. The spatiotemporal feature extraction module consists of a temporal feature extraction submodule and a spatial feature extraction submodule. The two submodules run in parallel and communicate with each other through a feature interaction interface. The temporal feature extraction submodule adopts a three-layer recurrent structure, with 128 hidden units in each layer. It processes historical load data and equipment operating status time-series data through a sliding window of size 24, capturing short-term dependencies and long-term evolution patterns of the data within a time scale of 1 hour to 72 hours. The spatial feature extraction submodule is based on a graph neural network architecture, dividing the power grid into 100 spatial nodes, each corresponding to a regional power supply unit. It constructs a spatial correlation matrix with a dimension of 100×100 by calculating the electrical and geographical distances between nodes, thereby mining the spatial coupling relationships and mutual influence characteristics of electricity loads in different regions. The output features of the two sub-modules are integrated collaboratively through a weighted fusion strategy. The fusion weights are dynamically allocated according to the data type. The weights for temporal features range from 0.4 to 0.6, and the weights for spatial features range from 0.4 to 0.6. The final output is a 256-dimensional spatiotemporal fusion feature vector, which lays the foundation for subsequent feature enhancement processing.

[0034] S3 utilizes a gated recurrent unit enhancement network to perform deep enhancement processing on the feature vector output by the spatiotemporal residual gated recurrent model, and strengthens the representation ability of the calibration load influence features by dynamically adjusting the gated mechanism parameters;

[0035] Specifically, step S3 initiates the gated recurrent unit enhancement network to perform deep enhancement processing on the 256-dimensional spatiotemporal fusion feature vector output from S2. This network includes a gating mechanism adjustment module, a feature enhancement module, and an output adjustment module. The gating mechanism adjustment module sets up two core gating structures: an update gate and a reset gate. Each gating structure is equipped with an independent parameter adjustment unit. The initial values ​​of the gating mechanism parameters are determined through a grid search algorithm. The initial weight matrix dimension of the update gate is 256×256, and the initial weight matrix dimension of the reset gate is also 256×256. The initial value of the bias term is set to 0.1. During processing, the gating mechanism adjustment module calculates the variance contribution value of each dimension of the feature vector in real time. Features with a variance contribution value greater than 0.05 are identified as key load-affecting features. For these features, the update gate weight is increased by 20% to 50%, and the reset gate weight is decreased by 10% to 30%, strengthening the transmission and retention of key features. For redundant features with a variance contribution value less than 0.01, their transmission is suppressed by the update gate to reduce interference from invalid information. The feature enhancement module uses a non-linear activation function to transform the selected features, and at the same time, it alleviates the gradient vanishing problem through a residual connection structure to ensure the effective transmission of features in the deep network. The final output is an enhanced feature vector with a dimension of 256 but a significantly improved ability to represent key features.

[0036] S4. Input the enhanced feature vector into the multimodal load prediction depth model, and construct the feature mapping relationship by combining the heterogeneity features of the multimodal data;

[0037] Specifically, step S4 inputs the enhanced feature vectors output from S3 into a multimodal load prediction deep model. This model includes a modality partitioning layer, a feature transformation layer, a fusion layer, and a mapping layer, with a dedicated processing flow designed for the heterogeneous features of multi-dimensional data. First, the modality partitioning layer divides the enhanced feature vectors into four sub-feature vectors based on data type: load modality, equipment modality, environment modality, and social modality. Each sub-feature vector corresponds to an independent feature transformation layer. The feature transformation layer adopts a fully connected network structure. The load modality feature transformation layer has two hidden layers, each with 128 neurons; the equipment and environment modality feature transformation layers each have two hidden layers, each with 64 neurons; and the social modality feature transformation layer has one hidden layer with 32 neurons. All feature transformation layers use batch normalization, with a normalization batch size of 32, mapping different modal features to a unified 128-dimensional feature space. Subsequently, the fusion layer calculates the association weights of various modal features through an attention mechanism. The weight calculation is based on the cosine similarity between features, with similarity values ​​ranging from -1 to 1. The association weight values ​​range from 0 to 1, and the sum of all modal weights is 1. Finally, a 128-dimensional multimodal fusion feature matrix is ​​generated through a weighted summation method. The mapping layer constructs a nonlinear mapping relationship between features and load prediction values ​​based on this matrix, providing accurate feature support for prediction calculation.

[0038] S5 generates preliminary load prediction results through the output layer of the multimodal load prediction deep model, and optimizes and adjusts the prediction results based on the feature weight allocation mechanism built into the model.

[0039] Specifically, step S5 calculates the mapping relationship constructed in S4 through the output layer of the multimodal load forecasting deep model, generating preliminary load forecast results. These results include load forecasts for the next 15 minutes, 30 minutes, 1 hour, and 24 hours, with an accuracy of 0.01 kW. The model's built-in feature weight allocation mechanism is based on a random forest algorithm to construct a feature importance evaluation model. This model consists of 100 decision trees, each with a depth of 10. Weights are determined by calculating the contribution of each feature dimension to the forecast results, using the Gini coefficient reduction as an indicator. Weight values ​​range from 0 to 0.02, with a sum of 1. The preliminary forecast results are then optimized based on these weights. Predictive components corresponding to core features with weights greater than 0.01 are strengthened to increase their impact on the final result; predictive components corresponding to secondary features with weights less than 0.005 are moderately suppressed to reduce noise interference. Meanwhile, the model sets rules to verify the reasonableness of the prediction results, compares the predicted values ​​with the historical maximum, minimum and average load values ​​for the same period, removes abnormal prediction data that exceed the reasonable range, ensures the reliability and reasonableness of the optimized prediction results, and finally outputs a preliminary load prediction dataset after weight adjustment and anomaly removal.

[0040] S6 verifies and outputs the optimized load forecast results through the smart grid data fusion computing platform, forming the final grid load forecast data.

[0041] Specifically, step S6 performs multi-dimensional data verification on the preliminary load forecast dataset optimized in S5 through the result verification module of the smart grid data fusion computing platform. The verification module includes three sub-modules: data integrity verification, logical consistency verification, and accuracy deviation verification. Data integrity verification checks whether the forecast dataset includes all forecast values ​​at the four time scales. If the missing rate exceeds 5%, a data completion mechanism is triggered, and missing data is supplemented through interpolation algorithms. Logical consistency verification compares the changing trends of forecast values ​​at different time scales to ensure that the growth or decline trend of forecast values ​​from 15 minutes to 24 hours conforms to the power grid load change pattern. If the trend conflict rate exceeds 10%, the model is returned for recalculation. Accuracy deviation verification uses two indicators: mean absolute error and root mean square error. The forecast value is compared with the actual load value of the same period in the most recent 30 days. The allowable range for mean absolute error is 0 to 5%, and the allowable range for root mean square error is 0 to 8%. For forecast results exceeding the allowable range, the model parameter fine-tuning process will be initiated. The verified forecast data is transmitted to the platform output module. The output module supports interface docking with six types of external systems, including the power grid dispatching system and the energy management platform. The final power grid load forecast data is output in real time through a standardized data format. The output data includes the forecast value, the forecast confidence interval, and the characteristic contribution analysis report. The confidence level of the forecast confidence interval is set to 95%, providing comprehensive and reliable data support for power grid dispatching decisions.

[0042] Preferably, the expression for the spatiotemporal residual gated cyclic model is: ,in, The feature output of the spatial location of s at time t. It is the sigmoid activation function. This is an element-wise multiplication operation. This is the model weight matrix. For bias terms, The input data is the spatial location of s at time t. For residual link terms, for Historical characteristics of spatial location at time s For time t Neighborhood characteristics of spatial location.

[0043] Specifically, the spatiotemporal residual gated recurrent model is designed for the collaborative capture of spatiotemporal correlation features in power grid load data. It achieves deep fusion of time-series dependencies and spatial distribution correlations through multi-dimensional parameter collaborative computation. During implementation, the model first determines the time step and spatial node partitioning scheme based on standardized data output from the smart grid data fusion computing platform. The time step is set to 15 minutes, and the spatial nodes are divided into 100 based on the power grid supply area, with each spatial node corresponding to an independent feature calculation unit. The weight matrices in the model are initialized using the Xavier initialization method, with a uniform dimension of 256×256 and an initial bias term value of 0.01. Iterative optimization is performed using the gradient descent algorithm, with 1000 iterations and an initial learning rate of 0.001, decaying by 50% every 200 iterations. The residual connection terms directly transmit the original input features through shortcut paths, effectively alleviating the gradient vanishing problem in deep network training and ensuring the effective transmission of spatiotemporal features deep within the model. This model controls the feature transfer ratio through the sigmoid activation function, combines the hyperbolic tangent function to achieve nonlinear transformation of features, and implements a gating mechanism to screen features through element-wise multiplication. The final output feature vector includes both the load evolution law in the historical time dimension and the load correlation characteristics in different spatial regions, providing comprehensive and accurate spatiotemporal fusion feature support for subsequent feature enhancement processing.

[0044] Preferably, the expression for the gated recurrent unit enhancement network is:

[0045] ,

[0046] ,

[0047] ,

[0048] ,

[0049] in, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. The input feature vector at time t, for Hide your status at all times. As a load characteristic enhancement factor, For weight parameters, This is the bias parameter.

[0050] Specifically, the gated recurrent unit (GRU) augmentation network enhances the representation capability of key features of the power grid load by dynamically coordinating the update and reset gates to achieve precise selection and enhancement of effective features. During implementation, the network first receives a 256-dimensional feature vector output from the spatiotemporal residual gated recurrent model. Simultaneously, a load feature enhancement factor is introduced, calculated based on the feature variance contribution, ranging from 0.1 to 0.9. Features with a variance contribution greater than 0.05 have enhancement factor values ​​no lower than 0.6. All weight parameters in the network are initialized using the He initialization method. The weight matrix dimension is adaptively set according to the input and output feature dimensions. The weight matrix dimension for both the update and reset gates is 256×256, while the weight matrix dimension for candidate hidden state calculation is 256×128. The initial value of the bias parameter is uniformly set to 0.02. During training, batch gradient descent is used to optimize parameters. The batch size is set to 64, the number of iterations is 800, and the learning rate is adjusted using an exponential decay strategy, with an initial learning rate of 0.002 and a decay coefficient of 0.95. The update gate outputs a value between 0 and 1 through the sigmoid activation function to control the retention ratio of historical hidden states; the reset gate also adjusts the effective information extraction of the current input features through the sigmoid activation function; after the candidate hidden state is transformed by the hyperbolic tangent function, it works with the output of the update gate to generate the final hidden state, thereby strengthening the key load features and suppressing redundant information, and greatly improving the representation ability of the feature vector.

[0051] Preferably, the expression for the multimodal load prediction depth model is: ,in, This is the load forecast value. For the number of multimodal data types, The weighting coefficients for the m-th modal data are... For the multilayer perceptron corresponding to the m-th modality, For convolutional feature extraction operations, Let m be the feature matrix of the m-th mode. For attention mechanism computational operations, This is the fusion feature matrix for all modalities.

[0052] Specifically, the multimodal load forecasting deep model addresses the heterogeneity of multimodal power grid data by constructing a dedicated feature fusion and prediction computation framework to achieve accurate integration and efficient prediction of different data types. During implementation, the number of multimodal data types is first defined, including four modalities: load, equipment, environment, and society, with a corresponding modality count of 4. The multilayer perceptron structure for each modality is designed based on the complexity of the modal features. The multilayer perceptron for the load modality includes three hidden layers, with 256, 128, and 64 neurons per layer, respectively; the multilayer perceptrons for the other three modalities each include two hidden layers, with 128 and 64 neurons per layer, respectively. Convolutional feature extraction uses a 3×3 convolutional kernel with a stride of 1 and SAME padding to extract local features for each modality. The attention mechanism is calculated based on the cosine similarity algorithm to quantify the correlation strength between features of different modalities, with the correlation strength ranging from -1 to 1. Modal weight coefficients are adaptively learned during the training process, with an initial value of 0.25. During training, the coefficients are dynamically adjusted based on the contribution of each modality to the prediction result, ensuring the sum of the weight coefficients is 1. During model training, the cross-entropy loss function is used to optimize the parameters, with 1200 iterations and a learning rate of 0.001, decreasing by 30% every 300 iterations. Through convolutional feature extraction, attention mechanism correlation calculation, and multilayer perceptron mapping, deep fusion and nonlinear mapping of multimodal features are achieved, resulting in accurate output of power grid load predictions.

[0053] Preferably, the data fusion expression of the smart grid data fusion computing platform is:

[0054]

[0055] Feature data, The number of dimensions for data collection. The fusion weights for the k-th dimension data are... Data is collected for the k-th dimension. The fusion bias for the k-th dimension data, For feature splicing operations, This is a normalization operation.

[0056] Specifically, the data fusion mechanism of the smart grid data fusion computing platform integrates multi-dimensional data collected from the power grid. Through weighted summation and feature concatenation, it generates fused feature data in a unified format. During implementation, the number of data collection dimensions is first determined, including 28 dimensions such as historical load values, equipment operating status, regional electricity consumption characteristics, and environmental influencing factors. The fusion weights for each dimension are initially determined using the analytic hierarchy process (AHP), and then iteratively optimized through model training. The weight values ​​range from 0 to 0.1, ensuring that the sum of all dimension weights is 1. The fusion bias term is set based on the statistical results of the mean of each dimension's data, with a value range from -0.05 to 0.05. The feature concatenation operation concatenates the data from each dimension in the order of collection dimensions, forming an original feature vector of dimension 1×(28×64). The Softmax normalization operation maps the weighted summed feature values ​​to between 0 and 1, achieving uniform scale for each dimension's features. The platform's data fusion process adopts a parallel computing architecture, supporting the processing of 100,000 data records per second, with a fusion latency controlled within 50 milliseconds. By using softmax normalization and element-wise multiplication of features, the original feature information of each dimension of data is preserved, and the correlation characteristics of different dimensions of data are strengthened, generating fusion feature data with 256 dimensions, providing standardized and high-value input data for subsequent spatiotemporal feature extraction.

[0057] Preferably, the parameter optimization expression for the power grid load forecast is: ,in, This is the optimal set of model parameters. The parameters to be optimized in the model For the sample size, Let i be the actual load value of the i-th sample. Let i be the predicted load value for the i-th sample. The regularization coefficient is . This is an L2 regularization term.

[0058] Specifically, the power grid load forecasting parameter optimization mechanism determines the optimal parameter set of the model by synergistically optimizing the minimization of forecasting error and regularization constraints, balancing forecasting accuracy and model stability. During implementation, the sample size is determined based on the historical data storage capacity of the smart grid data fusion computing platform, selecting load data from the past 365 days as training samples, with a sample size of 35,040 (collected at 15-minute intervals). The model parameters to be optimized include the weight matrix, bias term, modal weight coefficients, etc., totaling approximately 500,000 parameters. The regularization coefficient is determined using cross-validation, ranging from 0.001 to 0.01, with an optimal value of 0.005 to suppress overfitting. The parameter optimization process employs the stochastic gradient descent algorithm, with an initial learning rate of 0.003, and uses a momentum term to accelerate convergence, with a momentum coefficient set to 0.9. The optimization objective function consists of two parts: the first part is the sum of squared errors between the predicted and actual load values, used to measure forecasting accuracy; the second part is the L2 regularization term, used to constrain the parameter size. During the iterative optimization process, the objective function value is calculated every 50 iterations. The iteration stops when the change in the objective function value is less than 1e-6 over 100 consecutive iterations, and the optimal model parameter set is output. This optimization mechanism ensures that the model can fully fit the changing patterns of load data while avoiding overfitting that leads to a decrease in generalization ability, significantly improving the stability and reliability of power grid load forecasting.

[0059] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, obtaining the feature vector output by the spatiotemporal residual gated recurrent model, dividing the vector into a time-related feature subset and a spatially related feature subset according to the feature dimension, and identifying the two subsets through a feature classifier; S32, inputting the identified feature subsets into the input layer of the gated recurrent unit enhancement network, and setting the initial gate parameters according to the variance and mean of the feature subsets through the dynamic gate initialization module; S33, starting the gated recurrent unit enhancement network based on the initial gate parameters, filtering and retaining historical feature information through the update gate, and filtering noise from the current input features through the reset gate; S34, performing nonlinear transformation on the filtered features through the hidden layer of the gated recurrent unit enhancement network, strengthening the correlation between features of different dimensions in combination with the feature interaction module, and outputting the enhanced feature vector.

[0060] Specifically, step S3, the gated recurrent unit enhancement network feature processing flow, employs a four-step subdivision implementation mechanism to ensure accurate enhancement of key load features. During implementation, S31 first acquires the 256-dimensional feature vector output from the spatiotemporal residual gated recurrent model, dividing it into time-related and space-related feature subsets based on feature dimension attributes. The time-related feature subset includes 128-dimensional features, and the space-related feature subset also includes 128-dimensional features. A decision tree-based feature classifier is used for category identification, with the classifier accuracy set to above 98%. S32 inputs the identified two feature subsets into the gated recurrent unit enhancement network input layer. The dynamic gating initialization module calculates the variance and mean of the feature subsets, with the variance ranging from 0.01 to 0.5 and the mean ranging from 0.1 to 0.8. Based on the calculation results, the initial weight parameters for the update and reset gates are set, with the initial weight matrix dimension being 256×256 and the initial bias term value being 0.02. S33 starts the network based on initial gating parameters. The update gate outputs a value between 0 and 1 using the sigmoid function, retaining historical feature information with a variance contribution greater than 0.05. The reset gate also uses the sigmoid function to filter out noise in the current input features with a variance contribution less than 0.01. The hidden layers of the S34 network adopt a 3-layer fully connected structure, with 256, 128, and 256 neurons in each layer. Nonlinear transformation is achieved through the ReLU activation function. The feature interaction module calculates the correlation coefficient between two feature subsets, with the correlation coefficient ranging from -1 to 1, strengthening the interaction between highly correlated features. The final output is still a 256-dimensional enhanced feature vector, improving the key feature representation capability by more than 40%.

[0061] Preferred, such as Figure 3 As shown, step S4 includes the following sub-steps: S41, modally partitioning the feature vectors output by the gated recurrent unit augmentation network, and assigning them to the corresponding modality processing branches according to the data source and feature type; S42, each modality processing branch maps the input features to a feature space of a unified dimension through a dedicated feature transformation layer, and adjusts the feature scale of different branches using a modality adaptation function; S43, inputting the features output by different modality processing branches into the fusion layer of the multimodal load prediction deep model, and calculating the correlation strength between different modality features through a cross-modal attention mechanism; S44, performing weighted fusion of different modality features based on the correlation strength to generate a multimodal fusion feature matrix, which serves as the input for subsequent calculations of the multimodal load prediction deep model.

[0062] Specifically, step S4, the feature processing flow of the multimodal load prediction deep model, is refined to achieve efficient fusion of multimodal features through four collaborative operations. S31 first receives the 256-dimensional enhanced feature vector output from the gated recurrent unit (GRU) augmentation network. Based on the data source and feature type, it is divided into four modal processing branches: load, equipment, environment, and social. Each branch corresponds to a 64-dimensional feature vector. S42, each modal processing branch is equipped with a dedicated feature transformation layer. The load modal feature transformation layer uses a two-layer fully connected network with 128 and 64 neurons per layer. The other three modal feature transformation layers use a one-layer fully connected network with 64 neurons. All feature transformation layers employ batch normalization with a batch size of 32, uniformly mapping different modal features to a 64-dimensional feature space. The modality adaptation function adjusts the scale based on the feature standard deviation, which ranges from 0.1 to 0.3. S43 inputs the 64-dimensional features output from each branch into the fusion layer. The cross-modal attention mechanism calculates the correlation strength between features of different modalities based on the cosine similarity algorithm, with a correlation strength threshold set to 0.5 to filter high-correlation feature pairs. S44 assigns fusion weights based on the correlation strength, with weight values ​​ranging from 0.1 to 0.4, and the sum of all modal weights is 1. A 128-dimensional multimodal fusion feature matrix is ​​generated through weighted summation. This matrix retains the core features of a single modality while integrating the correlation information of different modalities, providing high-value feature support for subsequent prediction calculations, and improving feature fusion efficiency by more than 35%.

[0063] Preferred, such as Figure 4 As shown, step S5 includes the following sub-steps: S51, inputting the multimodal fusion feature matrix into the fully connected layer of the multimodal load forecasting deep model, and converting the fusion features into a predicted value vector through matrix multiplication; S52, calling the model's built-in feature weight allocation mechanism, calculating feature weight values ​​according to the contribution of different features to load forecasting, and adjusting the predicted value vector by weight; S53, applying range constraints to the weighted predicted value vector through the model's output adjustment layer, and eliminating abnormal predicted values ​​that exceed the reasonable load range; S54, using a sliding window smoothing strategy to locally optimize the adjusted predicted values, eliminate prediction fluctuations, and generate preliminary load forecasting results.

[0064] Specifically, step S5, the multimodal load prediction deep model prediction optimization process, is further subdivided into four precise operations to generate reliable preliminary load prediction results. S51 inputs a 128-dimensional multimodal fusion feature matrix into a fully connected layer, which includes two hidden layers with 128 and 64 neurons respectively. Matrix multiplication is used to convert the fusion features into a predicted value vector with four time scales (15 minutes, 30 minutes, 1 hour, and 24 hours), with a vector dimension of 1×4. S52 invokes a feature weight allocation mechanism based on random forests, which includes 100 decision trees, each with a depth of 10. The contribution of each feature dimension is calculated using the Gini coefficient reduction, with contribution values ​​ranging from 0.001 to 0.02. Weights are assigned based on the contribution, with the sum of the weights being 1, to weight and adjust the predicted value vector, strengthening the influence of core features. The S53 output adjustment layer sets a reasonable load range, which is determined based on the maximum and minimum load values ​​for the same period over the past 365 days. The maximum value is 1.1 times the historical peak value, and the minimum value is 0.9 times the historical trough value. Abnormal predicted values ​​exceeding this range are eliminated, with an outlier elimination accuracy of 99%. The S54 layer employs a sliding window smoothing strategy with a window size of 5. It performs a local weighted average calculation on the adjusted predicted values ​​to eliminate short-term prediction fluctuations. The smoothing coefficient ranges from 0.1 to 0.3, ultimately generating preliminary load prediction results with an accuracy of 0.01kW. The prediction fluctuation amplitude is reduced by more than 25%, providing a high-quality data foundation for subsequent platform verification.

[0065] like Figure 5As shown, a deep learning-based power grid load forecasting system is applied to a deep learning-based power grid load forecasting method. The system includes: a multi-dimensional data acquisition and transmission unit, used to acquire power grid load-related data through sensors and data acquisition terminals deployed in the smart grid, convert the data into a standardized format, and transmit it to the smart grid data fusion and computing unit; a smart grid data fusion and computing unit, connected to the multi-dimensional data acquisition and transmission unit, receiving the standardized data and performing fusion processing to generate feature data in a unified format, while also receiving and verifying the prediction results; and a spatiotemporal residual gated cyclic feature extraction unit, connected to the smart grid data fusion and computing unit, performing spatiotemporal correlation on the fused feature data. Feature capture and output of preliminary feature vectors; Gated cyclic unit enhancement processing unit, connected to spatiotemporal residual gated cyclic feature extraction unit, receives the preliminary feature vectors and strengthens the calibration features through dynamic gating mechanism, outputting enhanced feature vectors; Multimodal load forecasting calculation unit, connected to gated cyclic unit enhancement processing unit, performs multimodal fusion and forecasting calculation on enhanced feature vectors to generate optimized load forecasting results; Prediction result output and feedback unit, connected to multimodal load forecasting calculation unit and smart grid data fusion calculation unit, receives the verified final prediction results, outputs them to the dispatch center through the power grid dispatching interface, and simultaneously feeds back prediction error data to different model units for parameter adjustment.

[0066] In this invention, different scalar and vector parameters can be computed collaboratively in the same formula. The core premise is to eliminate data structure differences through dimensional adaptation and standardization. Vector parameters involved in the formula, such as feature vectors at different times and spatial locations, as well as multimodal feature matrices, are all converted into vector forms of a unified dimension through dimensional matching of the weight matrix. For example, neighborhood features and historical time feature vectors at different spatial locations are converted into feature representations of the same dimension through corresponding weight matrix operations. Scalar parameters, such as bias terms and regularization coefficients, are extended to dimensions matching the vectors through a broadcast mechanism, allowing them to participate in vector operations without changing their numerical attributes, ensuring computational compatibility at the dimensional level. Simultaneously, normalization operations map parameters of different magnitudes to the same numerical range, avoiding computational imbalances caused by differences in numerical ranges. The compatibility design of the computational rules provides technical support for the fusion of different types of parameters. The weighted summation and element-wise multiplication operations used in the formula support both feature superposition and interaction between vectors, and allow scalars to adjust and correct the overall vector structure. For example, the multiplication of the weight matrix and vector parameters enables linear transformation of features and information integration; scalar bias terms correct the offset of vector operation results through addition; and scalar weight coefficients in the attention mechanism quantify the importance of different vector features. These operational logics do not distinguish between parameter types, but only perform calculations based on numerical values, enabling scalars and vectors to participate in collaborative operations according to unified rules, achieving deep information fusion.

[0067] The inherent physical correlation between various parameters provides a rational basis for their calculation in the same formula. All parameters revolve around the core objective of power grid load forecasting. Vector parameters, such as spatiotemporal feature vectors and multimodal feature matrices, directly reflect the different dimensions and evolutionary patterns of load changes. Scalar parameters, such as bias terms, regularization coefficients, and weighting coefficients, are used to correct model biases, suppress overfitting, and adjust feature contributions. The two are complementary in function and share the same goal. For example, spatiotemporal feature vectors capture dynamic load changes, scalar bias terms correct systematic errors in the model, and scalar weighting coefficients highlight the influence of key features. Through the operational logic in the formula, they form a synergistic effect, jointly serving accurate load forecasting calculations and ensuring that the fusion of different types of parameters has clear physical meaning and practical value.

[0068] This deep learning-based power grid load forecasting method and system achieves both accuracy and efficiency in load forecasting through multi-model collaboration and a closed-loop design throughout the entire process. Methodologically, it leverages a smart grid data fusion computing platform to comprehensively collect and deeply fuse multi-dimensional load-related data, ensuring the integrity of the data foundation. A spatiotemporal residual gated cyclic model collaboratively captures the temporal dependence and spatial distribution correlation features in the data, overcoming the limitations of traditional single-dimensional feature mining. A gated cyclic unit enhancement network strengthens the representation of key features through a dynamic gating mechanism. The multimodal load forecasting deep model specifically addresses data heterogeneity, and combined with features weight allocation, outlier removal, and smoothing optimization mechanisms, significantly improves the reliability of the forecast results. At the system level, six functional units are closely integrated and clearly defined, achieving fully automated operation from data acquisition and transmission, feature extraction, enhancement processing to forecast output feedback, balancing data processing efficiency and forecast accuracy.

[0069] This method and system address the problem that traditional methods fail to fully exploit the spatiotemporal collaborative characteristics of load. By leveraging a dedicated module in the spatiotemporal residual gated cyclic model, it achieves the collaborative capture of temporal and spatial features. The gated cyclic unit enhancement network further strengthens the feature representation capability, enabling a comprehensive understanding of the dynamic evolution of load changes. Addressing the shortcomings of existing technologies such as incomplete multimodal data fusion and singular parameter optimization, this method utilizes a dedicated fusion mechanism within the smart grid data fusion computing platform to handle the heterogeneity of multi-source data. The multimodal load forecasting deep model, combined with a cross-modal attention mechanism, achieves accurate feature fusion. Simultaneously, a dedicated parameter optimization strategy balances forecasting performance and stability, effectively resisting noise data interference and meeting the high-precision and high-reliability requirements of smart grid load forecasting.

[0070] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power grid load forecasting method based on deep learning, characterized in that, Includes the following steps: S1. Collect multi-dimensional load-related data during power grid operation through a smart grid data fusion computing platform. This data includes historical load values, power grid equipment operating status data, regional electricity consumption characteristic data, and environmental influencing factor data. S2. Input the collected multi-dimensional data into a spatiotemporal residual gated cyclic model. The model's spatiotemporal feature extraction module collaboratively captures the time-series dependencies and spatial distribution correlation features in the data. S3. Utilize a gated cyclic unit enhancement network to perform deep enhancement processing on the feature vector output by the spatiotemporal residual gated cyclic model. Dynamically adjust the gate mechanism parameters to strengthen the characterization capability of the calibrated load influence features. S4. Input the enhanced feature vector into a multimodal load prediction deep model. Combine the heterogeneity characteristics of the multimodal data to construct a feature mapping relationship. S5 generates preliminary load forecast results through the output layer of the multimodal load forecasting deep model, and optimizes and adjusts the forecast results based on the built-in feature weight allocation mechanism of the model; S6 verifies and outputs the optimized load forecast results through the smart grid data fusion computing platform to form the final power grid load forecast data.

2. The power grid load forecasting method based on deep learning according to claim 1, characterized in that, The expression for the spatiotemporal residual gated loop model is: ,in, The feature output of the spatial location of s at time t. It is the sigmoid activation function. This is an element-wise multiplication operation. This is the model weight matrix. For bias terms, The input data is the spatial location of s at time t. For residual link terms, for Historical characteristics of spatial location at time s For time t Neighborhood characteristics of spatial location.

3. The power grid load forecasting method based on deep learning according to claim 1, characterized in that, The expression for the gated recurrent unit augmentation network is: ， ， ， ， in, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. The input feature vector at time t, for Hide your status at all times. As a load characteristic enhancement factor, For weight parameters, This is the bias parameter.

4. The deep learning-based power grid load forecasting method according to claim 1, characterized in that, The expression for the multimodal load prediction depth model is: ,in, This is the load forecast value. For the number of multimodal data types, The weighting coefficients for the m-th modal data are... For the multilayer perceptron corresponding to the m-th modality, For convolutional feature extraction operations, Let m be the feature matrix of the m-th mode. For attention mechanism computational operations, This is the fusion feature matrix for all modalities.

5. The deep learning-based power grid load forecasting method according to claim 1, characterized in that, The data fusion expression of the smart grid data fusion computing platform is: ； Feature data, The number of dimensions for data collection. The fusion weights for the k-th dimension data are... Data is collected for the k-th dimension. The fusion bias for the k-th dimension data, For feature splicing operations, This is a normalization operation.

6. The deep learning-based power grid load forecasting method according to claim 1, characterized in that, The parameter optimization expression for the power grid load forecast is as follows: ,in, This is the optimal set of model parameters. The parameters to be optimized in the model For the sample size, Let i be the actual load value of the i-th sample. Let i be the predicted load value for the i-th sample. The regularization coefficient is . This is an L2 regularization term.

7. The deep learning-based power grid load forecasting method according to claim 1, characterized in that, S3 includes the following sub-steps: S31, obtaining the feature vector output by the spatiotemporal residual gated recurrent model, dividing the vector into time-related feature subsets and spatially related feature subsets according to the feature dimension, and identifying the two subsets through a feature classifier; S32, inputting the identified feature subsets into the input layer of the gated recurrent unit enhancement network, and setting the initial gate parameters according to the variance and mean of the feature subsets through the dynamic gate initialization module; S33, starting the gated recurrent unit enhancement network based on the initial gate parameters, filtering and retaining historical feature information through the update gate, and filtering noise from the current input features through the reset gate; S34, performing nonlinear transformation on the filtered features through the hidden layer of the gated recurrent unit enhancement network, strengthening the correlation between features of different dimensions in combination with the feature interaction module, and outputting the enhanced feature vector.

8. The deep learning-based power grid load forecasting method according to claim 1, characterized in that, S4 includes the following sub-steps: S41, performing modal partitioning on the feature vectors output by the gated recurrent unit enhancement network, and assigning them to the corresponding modal processing branches according to the data source and feature type; S42, each modality processing branch maps the input features to a feature space of the same dimension through a dedicated feature transformation layer, and uses a modality adaptation function to adjust the feature scale of different branches; S43, input the features output from different modal processing branches into the fusion layer of the multimodal load prediction depth model, and calculate the correlation strength between different modal features through the cross-modal attention mechanism; S44, perform weighted fusion of different modal features based on the correlation strength to generate a multimodal fusion feature matrix, which serves as the input for subsequent calculations of the multimodal load prediction depth model.

9. The power grid load forecasting method based on deep learning according to claim 1, characterized in that, S5 includes the following sub-steps: S51, inputting the multimodal fusion feature matrix into the fully connected layer of the multimodal load forecasting deep model, and converting the fusion features into a predicted value vector through matrix multiplication; S52, calling the model's built-in feature weight allocation mechanism, calculating feature weight values ​​according to the contribution of different features to load forecasting, and adjusting the predicted value vector by weight; S53, applying range constraints to the weighted predicted value vector through the model's output adjustment layer, and eliminating abnormal predicted values ​​that exceed the reasonable load range; S54, using a sliding window smoothing strategy to locally optimize the adjusted predicted values, eliminate prediction fluctuations, and generate preliminary load forecasting results.

10. A power grid load forecasting system based on deep learning, characterized in that, This system, applied to the deep learning-based power grid load forecasting method described in claim 1, comprises: a multi-dimensional data acquisition and transmission unit, used to acquire power grid load-related data through sensors and data acquisition terminals deployed in the smart grid, convert the data into a standardized format, and transmit it to the smart grid data fusion and computing unit; a smart grid data fusion and computing unit, connected to the multi-dimensional data acquisition and transmission unit, receiving the standardized data and performing fusion processing to generate feature data in a unified format, while simultaneously receiving and verifying the prediction results; and a spatiotemporal residual gated cyclic feature extraction unit, connected to the smart grid data fusion and computing unit, for capturing spatiotemporal correlation features of the fused feature data. The system outputs a preliminary feature vector; a gated cyclic unit enhancement processing unit, connected to the spatiotemporal residual gated cyclic feature extraction unit, receives the preliminary feature vector and strengthens the calibration features through a dynamic gating mechanism, outputting an enhanced feature vector; a multimodal load forecasting calculation unit, connected to the gated cyclic unit enhancement processing unit, performs multimodal fusion and forecasting calculations on the enhanced feature vector, generating optimized load forecasting results; and a forecasting result output and feedback unit, connected to the multimodal load forecasting calculation unit and the smart grid data fusion calculation unit, receives the verified final forecasting results, outputs them to the dispatch center through the power grid dispatching interface, and simultaneously feeds back the forecasting error data to different model units for parameter adjustment.

Images