Multi-factor spatio-temporal power load prediction method and system based on deep learning

Abstract

The application discloses a multi-factor space-time power load prediction method and system based on deep learning, which comprises the following steps: collecting and cleaning multi-source heterogeneous data, constructing a space-time load data matrix and an external factor matrix; performing variational mode decomposition on the load sequence of each power grid spatial distribution area in the space-time load data matrix, and dividing the load sequence into a trend sequence, a seasonal sequence and a residual sequence; constructing a trend component model, a seasonal component model and a residual component model; using a physical perception gate unit downstream of each component model to dynamically adjust the contribution degree of each component model, and outputting a fused total power load prediction value; training each component model, fixing the parameters of each component model after the training is completed, optimizing and training a fusion model composed of each component model and the physical perception gate unit, and using the trained fusion model to perform power load prediction; and the application has the advantages of strong robustness and feasibility in engineering application.

Description

Technical Field

[0001] This invention relates to the field of power system automation and artificial intelligence, specifically to a multi-factor spatiotemporal power load prediction method and system based on deep learning. Background Technology

[0002] Power load forecasting is a core component of the safe and stable operation of a power system, and its accuracy directly impacts power generation planning, reserve capacity allocation, and economic benefits. Existing technologies are mainly divided into three categories:

[0003] 1. Traditional statistical methods: such as autoregressive integral moving average models and exponential smoothing. These methods assume that the data is stationary and have difficulty handling the nonlinear and non-stationary characteristics of load data (such as sudden fluctuations and complex seasonality), and the prediction error is usually high.

[0004] 2. Single deep learning models: such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), or Convolutional Neural Network (CNN). These models can capture time dependencies, but they have significant limitations: load includes multi-scale components such as trends, seasonality, and residuals, and a single model struggles to account for the morphological characteristics of different components (e.g., LSTM excels at trend prediction but is insufficient for fitting high-frequency seasonal fluctuations); multiple factors (such as temperature, humidity, and holidays) are usually simply pieced together, with a simplistic integration method that fails to distinguish their mechanisms of action on each component, easily leading to information redundancy or loss of effective information; spatiotemporal modeling capabilities are weak, often focusing on time series while ignoring inter-regional load correlations and spatial correlations caused by grid topology (e.g., the mutual influence of loads between regions).

[0005] 3. Decomposition-Prediction Fusion Method: In recent years, researchers have attempted to combine signal decomposition techniques (such as Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD)) with deep learning. For example, the VMD LSTM model decomposes the load into intrinsic mode functions (IMFs) and then uses LSTM for prediction, but it does not customize the model for the morphological differences of different IMFs, resulting in high prediction errors for seasonal components. Multiple factors are input into a single model without decoupling the correlation between factors and components. Graph Neural Networks (GNNs) handle spatial relationships, but without incorporating component decomposition, their robustness to random fluctuations is insufficient. A power load prediction method based on spatiotemporal graph convolutional networks under extreme weather conditions, disclosed in Chinese Patent Publication No. CN120822204A, employs a decomposition-prediction fusion method.

[0006] Summary of existing technical issues: 1. Distortion in component morphology modeling: The decomposed components of the load have significantly different statistical characteristics, but existing models adopt a unified architecture for smooth trends, strong periodic fluctuations and high-frequency random residuals, lacking specific physical evolution constraints, resulting in poor prediction robustness under extreme conditions.

[0007] 2. Lack of spatiotemporal dynamic correlation: The spatial constraints of the power grid topology on load distribution are ignored, and the power transfer and coordination characteristics between regions cannot be effectively captured, resulting in poor prediction robustness.

[0008] 3. Weak physical information calibration mechanism: Most existing methods are purely data-driven predictions and do not use physical information such as grid topology constraints and node power balance to constrain and correct the prediction results, which may lead to insufficient feasibility of the results in engineering applications.

[0009] 4. Simple fusion strategy: Most fusion strategies are linear weighting, which do not consider the nonlinear interaction between components and cannot perceive the dynamic changes in the power grid operating conditions, resulting in poor prediction robustness. Summary of the Invention

[0010] The technical problem to be solved by this invention is that existing power load forecasting methods have poor forecasting robustness and insufficient feasibility in engineering applications.

[0011] This invention solves the above-mentioned technical problems through the following technical means: a multi-factor spatiotemporal power load forecasting method based on deep learning, comprising: S1. Collect and clean multi-source heterogeneous data to construct a spatiotemporal load data matrix and an external factor matrix; the spatiotemporal load data matrix consists of load sequences of multiple power grid spatial distribution areas arranged according to time steps, and the external factor matrix consists of multiple external factors arranged according to time steps, including meteorological data, economic and social data, and dispatch and equipment information. S2. Perform variational mode decomposition on the load sequence of each power grid spatial distribution area in the spatiotemporal load data matrix, and divide it into trend sequence, seasonal sequence and residual sequence; S3. Match the trend series, seasonal series and residual series with external factors as input data, respectively, and input them into the trend component model, seasonal component model and residual component model built based on deep learning, respectively, and output the trend forecast value, seasonal load forecast value and residual forecast mean. S4. Downstream of the trend component model, seasonal component model and residual component model, a physical-confidence augmented matrix containing power grid topology information is constructed using a physical sensing gating unit. The physical-confidence augmented matrix is ​​then used to generate physical gating weights, dynamically adjust the contribution of each component model, and output the fused total power load forecast. S5. Train each component model, solidify the parameters of each component model after training, optimize and train the fusion model composed of each component model and the physical sensing gating unit to obtain the trained fusion model, and use the trained fusion model to predict power load.

[0012] Further, S2 includes: Determine the initial value of the total number of modes K1 and the upper limit of the number of modes, and perform variational mode decomposition on the load sequence of each power grid spatial distribution area to construct the following constrained optimization problem:

[0013] in, For the first One modal function, For the first The center frequency of each mode For Dirac functions, Denotes the square of the L2 norm. Indicates time The first-order partial derivative operator, Indicates the current moment. Represents the imaginary number symbol; Based on the current iteration round and Calculate the center frequency difference between adjacent modes ,like If it is greater than the threshold, let Proceed to the next round of optimization; if Stop iteration and output if the value is less than the threshold. As the optimal mode ; The load sequence of each power grid spatial distribution area is obtained by using variational mode decomposition. Decomposed into One mode; when and The corresponding modes are divided into trend components; the trend components are then sorted in chronological order to form a trend sequence. The sample entropy of the modality; when and Meanwhile, the period of the modal data is 24 hours or 168 hours, and the corresponding modality is divided into seasonal components; the seasonal components are sorted in chronological order to form a seasonal sequence. when and The corresponding modes are divided into residual components; the residual components are sorted in chronological order to form a residual sequence.

[0014] Furthermore, the construction process of the trend component model is as follows: The first input data is constructed by concatenating the trend sequence within a preset time window with long-term factors within that time window, namely GDP and weekly average temperature. The first input data and the spatial adjacency matrix are processed by a first graph convolutional network, outputting a spatial feature vector sequence. This spatial feature vector sequence is then fed into a 3-layer Transformer encoder to output the latent variables at the current time step. The similarity between latent variables is calculated using the following radial basis functions:

[0015] in, It is the first The implicit variables at time, express and similarity, For feature scale parameters, Denotes the square of the L2 norm. The signal variance of the trend series. The noise variance of the trend series; The Kronecker function; Let the feature matrix of the historical training set be... The corresponding true trend label vector is Construct the covariance matrix among historical training data based on the smoothing kernel function:

[0016] in, It is the identity matrix; Build and Covariance vector between ; future Predicted mean of the time-trend component model The calculation formula is:

[0017] in, It is a mean function; future Predictive dynamic variance of the time-trend component model The calculation formula is:

[0018] The final trend component model outputs the future. Hourly trend forecast value and its corresponding dynamic variance .

[0019] Furthermore, the construction process of the seasonal component model is as follows: The second input data is constructed by concatenating periodic factors with the same frequency within a preset time window, using a 1-hour step size, for the seasonal series. These periodic factors are temperature, humidity, and irradiance. The second input data, along with the spatial adjacency matrix, is then fed into a second graph convolutional network to extract spatial features. These spatial features are then fed into a two-layer sequentially connected one-dimensional convolutional network to extract local fluctuation features, followed by a single-layer LSTM to output the future... Hourly seasonal load forecast and its corresponding static error , This represents the historical mean squared error of the seasonal component model on the validation set.

[0020] Furthermore, the construction process of the residual component model is as follows: Gaussian noise is injected into the residual sequence, which is then fused with random factors and fed into a third-graph convolutional network along with the spatial adjacency matrix. The random factors include sudden changes in wind speed, holiday markers, and maintenance status. The spatial feature sequence output from the third-graph convolutional network is directly fed into the GRU, and the hidden state at the GRU's terminal is mapped to a residual feature vector. ; Calculate the residual eigenvectors at different times The similarity between them is calculated using the following formula:

[0021] in, It is the total number of mixed components; It is the first The mixed weights of the components; It is the first The covariance matrix of each component; It is the residual eigenvector at time i. yes and similarity, It is the first The center frequency vector of each component; Calculate the dynamic variance at the current prediction time. ,in, This is a Gram matrix constructed based on historical training data, where the element in the i-th row and j-th column is... ; yes The covariance vector between the training data and historical training data The variance of Gaussian noise is represented; future Hourly residual forecast mean The calculation formula is ,in, It is the residual true label vector in the historical training data; Finally, the residual component model outputs the future... Hourly residual forecast mean and its corresponding dynamic variance .

[0022] Furthermore, the spatial adjacency matrix construction process is as follows: An initial adjacency matrix is ​​established based on the power grid topology. ; Calculate the Pearson correlation coefficient matrix between load sequences in different spatial distribution areas of the power grid based on historical spatiotemporal load data. Pearson correlation coefficient matrix The Line number Column elements Representative area Load time series and regional loads within the historical observation window The degree of correlation of load time series within the same historical observation window; Final spatial adjacency matrix Obtained through weighted fusion: ,in, It is a balancing factor.

[0023] Furthermore, the step of constructing a physical-confidence augmented matrix containing power grid topology information using a physical sensing gating unit, generating physical gating weights using the physical-confidence augmented matrix, dynamically adjusting the contribution of each component model, and outputting the fused total power load prediction value includes: The physical sensing gating unit inputs a physical-confidence augmented matrix. To dynamically adjust the contribution of each component model, among which, The reactive power compensation state vector represents the real-time reactive power compensation device capacity data for each region's bus node. This is the capacity threshold vector, where represents the upper limit of the rated capacity of transformers in each region; Based on the constructed physical-confidence augmented matrix The formula for generating physical gating weights is:

[0024] in, For dynamic physical weight vectors, These correspond to the gating coefficients for the trend component, seasonal component, and residual component, respectively. Both are weight matrices. Both are bias vectors. for Activation function; The combined total power load forecast is ,in, This indicates dot product.

[0025] Furthermore, the training of each component model includes: The Adam optimizer was used with a learning rate of 0.001 and a batch size of 64. Each component model was trained in parallel on a GPU cluster. The optimization objective was to minimize the component loss function.

[0026] in, For the trend component loss, negative log-likelihood loss is used; To account for residual component loss, maximum likelihood estimation or variational inference of the objective is employed; For seasonal component loss and , These are the weighting coefficients. The first data-driven term is defined by the following formula:

[0027] in, This refers to the training batch size; The time window length of the seasonal series; For the first The true seasonal load components at time t obtained from each training batch through variational mode decomposition; For the first The seasonal load forecast output at time t of the seasonal component model in each training batch; The first physical residual term is defined by the following formula:

[0028] in, This represents the seasonal load forecast output by the seasonal component model at time t. This represents the total number of time steps involved in the physical loss calculation. Equivalent indoor temperature Outdoor temperature For energy efficiency ratio, For equivalent thermal resistance, For equivalent heat capacity, This represents the square of the L2 norm.

[0029] Furthermore, the optimization training of the fusion model composed of each component model and the physical perception gating unit includes: After training the three component models, the parameters of all component models are fixed, and the learnable parameters of the physical perception gating unit are iteratively adjusted. The value of the overall optimization objective function is calculated during each iteration until the value of the overall optimization objective function is minimized or the preset number of iterations is reached. Then, the iteration stops, and the trained fusion model is obtained. The formula is as follows:

[0030] in, and For balancing weighting coefficients; As the second data-driven term, the weighted mean square error is used to calculate the fused total power load forecast. Compared with the actual load value Deviation; The second physical residual term is defined by the following formula:

[0031] in, For the number of regions, To predict the time window, For a moment Total power supply For the estimated network loss value, For the region exist The actual load value at any given moment.

[0032] This invention also provides a system employing the above-described deep learning-based multi-factor spatiotemporal power load forecasting method, comprising: The data acquisition module is used to collect and clean multi-source heterogeneous data, and construct a spatiotemporal load data matrix and an external factor matrix. The spatiotemporal load data matrix consists of load sequences of multiple power grid spatial distribution areas arranged according to time steps, and the external factor matrix consists of multiple external factors arranged according to time steps. The external factors include meteorological data, economic and social data, and dispatch and equipment information. The component partitioning module is used to perform variational mode decomposition on the load sequence of each power grid spatial distribution area in the spatiotemporal load data matrix, dividing it into trend sequence, seasonal sequence and residual sequence. The component model construction module is used to match external factors as input data to the trend series, seasonal series and residual series respectively. The inputs are the trend component model, seasonal component model and residual component model built based on deep learning, and the outputs are the trend forecast value, seasonal load forecast value and residual forecast mean. The fusion output module is used downstream of the trend component model, seasonal component model and residual component model to construct a physical-confidence augmented matrix containing power grid topology information using physical sensing gating units, and to generate physical gating weights using the physical-confidence augmented matrix, dynamically adjust the contribution of each component model, and output the fused total power load forecast value. The model training module is used to train each component model. After training, the parameters of each component model are fixed. The fusion model composed of each component model and the physical sensing gating unit is optimized and trained to obtain the trained fusion model. The trained fusion model is then used for power load forecasting.

[0033] The advantages of this invention are: 1. This invention decomposes load data into components with well-defined morphological characteristics and, based on the statistical distribution characteristics of each component, custom-constructs several prediction architectures, including trend component models, seasonal component models, and residual component models, to achieve accurate fitting of different load components and strong prediction robustness. In the fusion stage, the contribution of each component model is dynamically adjusted by embedding a physical sensing gating unit into the power grid topology, taking into account actual physical information, making the model highly feasible in engineering applications. Furthermore, based on customized component modeling, the introduction of physical priors enhances the prediction stability under complex operating conditions, resulting in strong robustness.

[0034] 2. This invention realizes a physical-data dual-driven spatiotemporal prediction architecture, which explicitly embeds a physical consistency kernel function (radial basis function) and thermodynamic partial differential equations into the component model. This architecture uses physical mechanisms to dynamically constrain the neural network, effectively improving the prediction ability of existing models under extreme weather or data sparse conditions.

[0035] 3. This invention constructs a multi-factor coupled component-customized model, and customizes a three-in-one modeling strategy of "factor-model-constraint" for different load components. The trend component model integrates economic indicators and manifold smoothing constraints, the seasonal component model integrates meteorological factors and thermodynamic equation constraints, and the residual component model integrates sudden events and Bayesian uncertainty quantification.

[0036] 4. This invention proposes a dynamic weighted intelligent fusion strategy based on physical perception. The fusion weights are dynamically generated according to the real-time operating conditions of the power grid and the prediction variance of each component, rather than using traditional linear weighting. By introducing physical criteria such as Kirchhoff power balance as auxiliary constraints to optimize the hybrid loss function, the prediction results not only achieve statistical optimality but also satisfy physical logic consistency, significantly enhancing the system's robustness in extreme scenarios. Attached Figure Description

[0037] Figure 1This is an overall flowchart of the deep learning-based multi-factor spatiotemporal power load forecasting method disclosed in the embodiments of the present invention; Figure 2 This is a diagram of the trend component model architecture in the deep learning-based multi-factor spatiotemporal power load forecasting method disclosed in the embodiments of the present invention. Figure 3 This is a diagram of the seasonal component model architecture in the deep learning-based multi-factor spatiotemporal power load forecasting method disclosed in the embodiments of the present invention. Figure 4 This is a diagram of the residual component model architecture in the deep learning-based multi-factor spatiotemporal power load prediction method disclosed in the embodiments of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1 like Figure 1 As shown, Embodiment 1 of the present invention provides a multi-factor spatiotemporal power load forecasting method based on deep learning, including the following steps: S1. Collect and clean multi-source heterogeneous data, including power grid operation load data and external factor data. External factor data includes meteorological data, economic and social data, and dispatch and equipment information. The content of each data item is as follows: Power grid operation load data: Load sequences (time resolution ≤ 1 hour) from multiple spatially distributed areas of the power grid (such as busbars, substations, or administrative regions) constitute a spatiotemporal load data matrix. , where N is the number of regions and T1 is the time step.

[0040] Meteorological data: temperature, humidity, wind speed, and irradiance, etc. Economic and social data: GDP, industrial capacity utilization rate, holidays, and user category statistics; Dispatch and equipment information: maintenance plans and regional power transfer records.

[0041] The aforementioned multi-source heterogeneous data undergoes preprocessing including time alignment, spatial gridding, missing data imputation, outlier removal, and normalization. The preprocessed spatiotemporal load data matrix is ​​then output. and external factor matrix , The quantity of data related to external factors.

[0042] S2. Load Decomposition: To address the non-stationarity of power load signals, this invention employs the Variational Mode Decomposition (VMD) algorithm. The initial value of the total number of modes K1 and the upper limit of the number of modes are determined, and the load sequence for each power grid spatial distribution area is analyzed. Variational mode decomposition is performed, and the following constrained optimization problem is constructed. By minimizing the sum of the estimated bandwidths of each mode, the non-stationary power load signal is decomposed into a stationary component with a specific center frequency, thereby determining the optimal number of modes:

[0043] in, For the first One modal function, For the first The center frequency of each mode For Dirac functions, Denotes the square of the L2 norm. Indicates the current moment. Indicates time The first-order partial derivative operator, Represents the imaginary number symbol.

[0044] Based on the current iteration round and Calculate the center frequency difference between adjacent modes .like If the value exceeds the threshold (0.05 Hz in this implementation), it indicates that there is still room for spectrum separation, and the system automatically... Proceed to the next round of full optimization; if If the value is less than the threshold, it is considered over-decomposition, the iteration is stopped, and the result is output. As the optimal mode In this implementation, the optimal number of modes is automatically determined using the above criteria. Therefore, variational mode decomposition is used to analyze the load sequence of each power grid spatial distribution area. Decomposed into One modality.

[0045] Using the sample entropy (SE) and center frequency of each mode As a quantitative criterion, based on sample statistical analysis Each modality is clustered into three classes to eliminate modality aliasing and adapt to subsequent customized modeling: Trend components: and Located in the lowest frequency band, it exhibits low-frequency smooth changes, reflecting the long-term drift and macroscopic evolution of power load; the trend components are sorted in chronological order to form a trend sequence.

[0046] Seasonal ingredients: and Meanwhile, the modal data exhibits significant periodicity in the time domain, with its dominant period being close to 24 hours or 168 hours, reflecting the daily and weekly fluctuation patterns driven by meteorological factors; the seasonal components are sorted in chronological order to form a seasonal sequence.

[0047] Residual components: and It is in the highest frequency band, exhibiting high-frequency random fluctuations, reflecting system noise, non-periodic sudden events and disturbances; the residual components are sorted in chronological order to form a residual sequence.

[0048] It should be noted that variational mode decomposition is an existing technology, and it does not improve the calculation of sample entropy, mode function, and center frequency. These will not be elaborated here. This invention only improves the optimal number of modes and the mode component partitioning method of variational mode decomposition.

[0049] S3. Customized Component Modeling: This step involves constructing deep prediction models, including Graph Convolutional Networks (GCNs) and physical mechanism constraints, for the trend, seasonal, and residual components obtained after VMD decomposition, based on their different physical and statistical characteristics. These models are the trend component model, seasonal component model, and residual component model, respectively, to accurately capture the spatiotemporal characteristics of multiple factors. The following provides a detailed introduction to each model.

[0050] (1) Trend component model: Transformer architecture based on Gaussian process smoothing kernel Input: The first input data is constructed by concatenating a trend sequence window (time window length 168 hours) with long-term factors (GDP, weekly average temperature) within that time window length.

[0051] Structure: such as Figure 2 As shown, a manifold-constrained spatiotemporal coupling architecture is constructed. First, the first input data and spatial adjacency matrix are processed by a first graph convolutional network, mapping discrete spatial nodes into a sequence of spatial feature vectors containing topological information. Then, this sequence of spatial feature vectors is directly fed as input to a 3-layer Transformer encoder (see...). Figure 2 The Transformer encoder (including multi-head attention, residual connections, and layer normalization operations) utilizes a self-attention mechanism to further extract long-term temporal evolution patterns based on spatial features, outputting latent variables in deep spatiotemporal space. ; Construction of a smoothing kernel output layer: To address the lack of physical smoothness in traditional Transformer outputs, this invention abandons the traditional fully connected regression layer and innovatively incorporates the latent variables of the Transformer output. The mapping is represented by the induction point of the Gaussian process. The similarity between latent variables is calculated using radial basis functions, and the covariance matrix is ​​constructed as follows:

[0052] in, The output of the last layer of the Transformer encoder is the first The implicit variables at time, express and similarity, This is a characteristic scale parameter used to control the smooth decay rate of trend changes; The signal variance of the trend sequence determines the potential oscillation amplitude of the trend sequence; The noise variance of the trend series; For standard Kronecker functions, their subscripts are... and Consistent with the aforementioned meaning of time step, if and only if When the condition is met, the function takes the value 1; otherwise, it takes the value 0.

[0053] This results in a prediction that follows a multivariate Gaussian distribution.

[0054]

[0055] in, It is a Gaussian process, meaning that the predicted value of the trend sequence is not a single value, but a set of stochastic processes. This Gaussian process is composed of the mean function. and radial basis functions Uniquely certain; yes The mean function represents the trend component model in The mathematical expectation of the probability prediction distribution given at a single moment.

[0056] To obtain accurate prediction results from the aforementioned multivariate Gaussian distribution, this invention, based on the posterior inference principle of Gaussian process regression, introduces the feature matrix of the historical training set and its corresponding true trend labels to calculate the predicted mean and dynamic variance at the target time. The specific calculation steps are as follows: Let the feature matrix of the historical training set be... Its elements are a set of deep spatiotemporal feature vectors extracted from samples in the historical training set through a Transformer layer. The corresponding true trend label vector is... Constructing the covariance matrix among historical training data based on a smoothing kernel function:

[0057] in It is an identity matrix.

[0058] and hidden variables With historical feature matrix Covariance vector between

[0059] future Forecast mean of time trend component The calculation formula is:

[0060] future Predictive dynamic variance of time-trend components The calculation formula is:

[0061] Output: Future Hourly trend forecast value and its corresponding dynamic variance This variance is calculated in real time by the Gaussian process regression layer based on the distribution characteristics of the latent variables, and is used to quantify the potential drift risk of long-term evolution in each region.

[0062] This model can capture long-term time-series dependencies across regions and is supplemented with a Gaussian process smoothing kernel to implement physical constraints, thus solving the problems of insufficient learning of long-term dependencies and non-physical trend shifts in traditional LSTM.

[0063] It should be noted that, in order to ensure the smoothness of the trend evolution manifold and suppress non-physical drift, in the subsequent component model training phase, the network weights and Gaussian smoothing kernel parameters in this architecture will be adjusted using the trend component loss. A joint optimization solution is performed, and the trend component loss adopts negative log-likelihood loss. Negative log-likelihood loss is an existing technique, and the specific principle can be found in the literature "Rasmussen CE, Williams CK I. Gaussian Processes for Machine Learning[M].MIT press, 2006.".

[0064] (2) Seasonal component model: Physical information driven graph convolution LSTM Input: Seasonal series (time window length 24 hours) and within this time window length, periodic factors (temperature, humidity and irradiance) with the same frequency are spliced ​​together in 1-hour increments to construct the second input data; Structure: such as Figure 3As shown, the second input data and the spatial adjacency matrix are input into the second graph convolutional network to extract spatial features. The spatial features are then input into a two-layer one-dimensional convolutional network (CNN with 3 kernels and 64 channels) connected in sequence to extract local fluctuation features. Subsequently, a single-layer LSTM (128 units) is used for time-dependent modeling. Physical constraint embedding mechanism: The thermodynamic energy conservation equation is embedded into the training loss function of the seasonal component model. Specifically, the output value of the seasonal component model... The variable to be solved in the thermodynamic energy conservation equation .

[0065] Thermodynamic energy conservation equation:

[0066] in, Equivalent indoor temperature Outdoor temperature For seasonal loads predicted by the seasonal component model, For energy efficiency ratio, For equivalent thermal resistance, This is the equivalent heat capacity.

[0067] It should be noted that, The hidden state variables constructed for the model are not directly collected observation values. During training, they are automatically deduced using the physical constraints of PINN (Physically Informed Neural Network). The trajectory of change, In this embodiment, the trainable parameters of the model are assigned physical experience values ​​during initialization. During training, the backpropagation algorithm is used to minimize the physical loss function, eliminating the need for manual on-site measurement.

[0068] During the backpropagation training phase, it is calculated whether the network output value satisfies the above thermodynamic energy conservation equation (i.e., the physical residual is calculated). If it does not satisfy the equation, then the first physical residual term is... The data is then sent back to the network, forcing the seasonal component model to learn weight parameters that conform to thermodynamic physical laws.

[0069] Define seasonal component loss Among them, the first data-driven item The formula used to supervise the model's fitting accuracy to historical data is as follows:

[0070] in, This refers to the training batch size; The time window length of the seasonal series; For the first The true seasonal load components at time t obtained from each training batch through variational mode decomposition; For the first The seasonal load predictions output by the seasonal component model at time t in each training batch.

[0071] First physical residual term The formula used to quantify the deviation of the prediction results from the thermodynamic law of conservation of energy is as follows:

[0072] in, This represents the seasonal load forecast output by the seasonal component model at time t. This represents the total number of time steps involved in the physical loss calculation.

[0073] By minimizing seasonal component loss This ensures that the seasonal component model's predictions under extreme weather conditions conform to the physical heat transfer mechanism.

[0074] Output: Future Hourly seasonal load forecast and its corresponding static error .in, The historical mean squared error of the seasonal component model on the validation set is used as a static constant output to characterize the basic reliability of the seasonal component model.

[0075] The seasonal component model can accurately map meteorological response patterns using a physical information neural network embedded with the thermodynamic energy conservation equation, solving the problems of poor generalization and lack of physical interpretability of traditional black-box models under extreme weather conditions.

[0076] (3) Residual component model: Spatiotemporal deep kernel learning GCN-GRU probabilistic prediction architecture Input: residual sequence (time window length 48 hours) and random factors (wind speed mutation, holiday markers, and maintenance status) corresponding to the time window length; among them, holiday markers and maintenance status are converted into unique heat codes and aligned to the corresponding time window by hourly steps to capture non-periodic sudden load fluctuations.

[0077] Structure: such as Figure 4 As shown, a deep feature extraction architecture is adopted by cascading sparse GCN and GRU.

[0078] First, Gaussian noise is injected into the residual sequence to enhance the robustness of the model. Then, it is fused with random factors and fed into a third graph convolutional network along with the spatial adjacency matrix. Spatial convolution is performed in combination with the spatial adjacency matrix to extract the spatial propagation features of the residual signal on the power grid topology. The spatial feature sequence output by the graph convolutional network is directly fed into the GRU (with 32 units and a Dropout mechanism) to capture the decay and abrupt change characteristics of the residual signal in the time dimension. The terminal hidden states of GRU are mapped to a low-dimensional residual feature vector. It is important to emphasize that this feature vector Instead of being used directly as the final predicted value, they are used as the input coordinates for subsequent spectral mixing kernel functions to construct the feature space of the Gaussian process.

[0079] Spectral Mixture Kernel Probabilistic Modeling: Define the spectral mixture kernel function (SM kernel) to compute the residual eigenvectors at different time points. The similarity between them. This similarity measure is not only based on Euclidean distance, but also on the degree of matching of frequency components, thereby capturing the complex quasi-periodic patterns hidden in the residuals. The formula is as follows:

[0080] in, It represents the total number of mixed components, indicating the number of different frequency components contained in the residual; It is the first The mixed weights of the components; It is the first The covariance matrix of each component controls the variation of that component with distance. The rate of increase and decrease; It is the residual eigenvector at time i. yes and similarity, It is the first The center frequency vectors of each component are used to capture the main periodic patterns in the residual signal. The parameters mentioned above { , , All of these are trainable parameters.

[0081] Deriving the dynamic uncertainty variance, based on the analytical solution of Gaussian process regression, and using the aforementioned kernel function to calculate the dynamic variance at the current prediction time. :

[0082] in, It is the prior variance (autocorrelation) calculated by the kernel function; This is a Gram matrix constructed based on historical training data, where the element in the i-th row and j-th column is... This matrix stores the spatiotemporal correlation structure among all samples in the historical residual sequence; yes The covariance vector between the training data and historical training data This represents the variance of Gaussian noise.

[0083] When encountering unseen sudden disturbances (such as extreme weather or equipment failure), the residual feature vector extracted by the network... It will deviate from the historical distribution, causing its covariance vector with historical data to differ. The numerical values ​​were significantly reduced, ensuring that the residual component model could issue high-risk warnings for abnormal operating conditions.

[0084] Finally, based on the posterior inference principle of Gaussian process regression, the predicted mean of the residual component model is calculated, and the conditional probability density function is finally constructed.

[0085] Calculate using kernel function Similarity to historical training data, thus inferring future... Hourly residual forecast mean The calculation formula is as follows:

[0086] in, It is the residual true label vector from the historical training data.

[0087] This formula indicates the current predicted value. Essentially, it is a weighted average of historical residual data, with the weights determined by the similarity between the current state and historical states. Decide.

[0088] Based on the mean and variance calculated above, the conditional probability density function of the residual components is constructed as follows:

[0089] in, These are the input features of the residual component model. Represents the true value, dynamic variance It reflects the intensity of the disturbance at the current moment.

[0090] Output: Future Hourly residual forecast mean and its corresponding dynamic variance This variance is derived analytically from the Gaussian process of the spectral mixing kernel function and is used to quantify in real time the predicted failure risk due to sudden disturbances or unseen operating conditions.

[0091] The residual component model can capture the implicit quasi-periodic features in the residuals through deep kernel learning and realize Bayesian uncertainty quantification, which solves the problem that high-frequency random disturbances are difficult to model accurately and the prediction risk is difficult to measure.

[0092] To accurately quantify the variance of spatiotemporal uncertainty, the internal parameters of this deep kernel learning architecture will be processed through residual component loss during the subsequent component model training phase. Iterative optimization was performed, and the residual component loss was evaluated using maximum likelihood estimation or variational inference. Maximum likelihood estimation or variational inference is an existing technique, and the specific principles and derivations can be found in the classic literature "Blei DM, Kucukelbir A, McAuliffe DD. Variational inference: A review for statisticians[J]. Journal of the American Statistical Association, 2017, 112(518): 859-877.".

[0093] It should be noted that, to fully utilize the spatial coupling relationship between power loads in different regions, this invention embeds a graph convolutional network layer in each component model. The same spatial adjacency matrix is ​​used for the trend component model, seasonal component model, and residual component model. The spatial adjacency matrix construction process is as follows: An initial adjacency matrix is ​​established based on the power grid topology (substation-bus-line connection relationship). ; Calculate the Pearson correlation coefficient matrix between load sequences in different spatial distribution areas of the power grid based on historical spatiotemporal load data. This represents the dynamic similarity of loads between regions. (Pearson correlation coefficient matrix) The Line number Column elements Representative area Load time series and regional loads within the historical observation window The degree of correlation of load time series within the same historical observation window; The final spatial adjacency matrix is ​​obtained through weighted fusion: ,in It is a balancing factor.

[0094] The spatial adjacency matrix performs the following spatial feature propagation mechanism at each layer of the GCN:

[0095] in, For the normalized adjacency matrix, For GCN's first Features of the spatial adjacency matrix of the layer input, For learnable weight matrix, This is the ReLU activation function.

[0096] This mechanism enables spatial information exchange and feature diffusion of regional loads, allowing the model to capture power transfer patterns and linkages between multiple nodes.

[0097] S4. Prediction result fusion, specifically based on gated fusion using physical information inductive bias, is implemented as follows: To address the issue of traditional fusion models' inability to adapt to complex physical condition switching, this embodiment employs a Physical Perception Gated Unit (PGU). This unit, serving as the system's fusion decision module, is located downstream of the trend, seasonal, and residual component models, establishing a logical interactive closed loop with each component model for feature feedback and weight assignment. This unit receives a physical-confidence augmented matrix as input. This allows us to perceive the reliability of each component model under the current physical conditions, thereby dynamically adjusting the contribution of each component model and achieving adaptive nonlinear calibration of the prediction results.

[0098] Physical-confidence augmented matrix The construction process is as follows: First, the original parameters reflecting the physical state of the power grid and the prediction uncertainty indicators of each component model are extracted in real time, including dynamic topology parameters and reactive power compensation status. The transformer physical capacity threshold c and the variance characteristics of the component model.

[0099] Dynamic topology parameters: directly using the initial adjacency matrix. It is used to characterize the physical connection relationship between busbars and lines between regions; Reactive power compensation state vector Extract the real-time reactive power compensation device (such as capacitors and reactors) input capacity data of each area bus node to construct a reactive power compensation state vector; Capacity threshold vector c: Extract the upper limit of the rated capacity of transformers in each region to construct a capacity threshold vector, which serves as the physical boundary constraint for load growth.

[0100] Component model variance characteristics: Extracting the predictive dynamic variance of the trend component models for each region. Static error of seasonal component model and the dynamic variance of the residual component model And construct the corresponding variance vector.

[0101] To eliminate the influence of different units on weight generation, the above parameters are linearly normalized so that their values ​​are mapped to... interval; With the initial adjacency matrix Based on the fundamental features, the normalized reactive power compensation state vector, capacity threshold vector, and variance vectors of each component model are horizontally concatenated according to the corresponding region nodes to construct the physical-confidence augmented matrix. This is used to characterize the current physical operating status of the power grid and the real-time confidence level of the model prediction. Different rows of the physical-confidence augmented matrix are different nodes, and the columns are feature splicing (topology, physical state, variance).

[0102] Based on the constructed physical-confidence augmented matrix The formula for generating physical gating weights is:

[0103] in, For dynamic physical weight vectors, These correspond to the gating coefficients for the trend component, seasonal component, and residual component, respectively. Both are weight matrices. Both are bias vectors. for Activation function.

[0104] The primary physics fusion output formula is:

[0105] in, Dot product, This represents the total predicted power load after integration. The structure utilizes physical priors... The intervention of [the system] makes the model structurally inclined to predict combinations that conform to physical common sense.

[0106] S5. Model Training and System Implementation; The system training process is divided into two stages: (1) Component model training phase: The component model training phase uses the Adam optimizer with a learning rate of 0.001 and a batch size of 64. Each component model is trained in parallel on a GPU cluster, with the optimization objective being to minimize the component synthesis loss function.

[0107] Among them, the trend component loss Negative log-likelihood loss is used to optimize the smoothing kernel parameter of the Gaussian process regression layer, ensuring the smoothness of the trend evolution manifold and suppressing non-physical drift. Seasonal component loss Defined as By introducing the thermodynamic energy conservation equation as a constraint, the model is forced to learn the mapping relationship between meteorological conditions and load that conforms to the physical heat transfer mechanism. Residual component loss The target is estimated using maximum likelihood estimation or variational inference, and the probability distribution of random perturbations is captured using spectral mixing kernels to quantify the spatiotemporal uncertainty variance. .

[0108] (2) Fusion model training phase: After pre-training the three component models, the fusion model training phase begins. The fusion model is an overall prediction architecture composed of the "trained three-component model with fixed parameters" and the "Physical Awareness Gated Unit (PGU) to be optimized." All component model parameters must be fixed to ensure that the predicted values ​​and confidence features input to the PGU remain statically stable, thereby eliminating gradient instability during training. The fusion model training phase employs a segmented learning rate decay strategy, decreasing by 0.9 every 50 epochs. The overall training scheme uses an early stopping mechanism (terminating training if the validation set loss increases for five consecutive epochs).

[0109] Iterative optimization of the learnable parameters within the PGU, specifically including: the parameters used to augment the physics-confidence matrix. Weight matrix mapped to weight distribution and bias vector In each iteration, the value of the overall optimization objective function is calculated until the value of the overall optimization objective function is minimized or the preset number of iterations is reached. Then, the iteration stops, and the trained fusion model is obtained. Through the penalty mechanism of the physical residual term, the error gradient is only backpropagated within the PGU and its weight allocation logic is corrected. This forces the model to perform nonlinear fusion correction within the physically feasible space, ultimately achieving alignment between physical correction and data patterns.

[0110] The above process optimizes the overall objective function. The design is as follows:

[0111] The overall optimization objective function is a loss function architecture that incorporates physical information and machine learning, ensuring the accuracy of the fused total power load forecast. It does not deviate from the basic physical laws of the power system (such as the conservation of energy). and To balance the weighting coefficients, the second data-driven term Calculate using weighted mean square error Compared with the actual load value The deviation. Second physical residual term. Based on the KCL power balance constraint, the formula is as follows:

[0112] in, For the number of regions, For the forecast time window (future H hours). For a moment Total power supply For the estimated network loss value, For the region exist The actual load value at any given moment.

[0113] By constraining the physical residual terms mentioned above, when the load prediction results generated by the deep learning model violate the physical power balance law, the system generates a large penalty value, thereby correcting the weight parameters and ensuring that the fusion result always operates within the physically feasible solution space. This method significantly improves the model's prediction robustness under special scenarios such as extreme weather or sudden load changes.

[0114] (3) System deployment: Once the model is trained, it can be exported to the prediction engine module and deployed in a Hadoop / Spark distributed environment. The prediction engine provides a real-time API call interface in the form of Python, supporting batch prediction and rolling window updates.

[0115] Through the above technical solutions, this invention decomposes load data into components with clearly defined morphological characteristics, and customizes a deep kernel learning, spatiotemporal feature extraction, and probabilistic prediction architecture based on the statistical distribution characteristics of each component, achieving accurate fitting of different load components. In the fusion stage, the prediction results are integrated through a spatiotemporal graph modeling strategy embedded with the power grid topology, and further physical perception gating and mechanistic constraints are introduced as advanced calibration mechanisms. This allows the model to improve prediction stability under complex operating conditions by introducing physical priors, based on component-customized modeling. This method emphasizes the "regularity of the morphology of components of the same type," significantly improving prediction accuracy and robustness, and is particularly suitable for short-term (24-72 hours) and medium-term (1-7 days) power load forecasting.

[0116] Example 2 Embodiment 2 of the present invention provides a system employing the method of Embodiment 1, comprising: The data acquisition module is used to collect and clean multi-source heterogeneous data, and construct a spatiotemporal load data matrix and an external factor matrix. The spatiotemporal load data matrix consists of load sequences of multiple power grid spatial distribution areas arranged according to time steps, and the external factor matrix consists of multiple external factors arranged according to time steps. The external factors include meteorological data, economic and social data, and dispatch and equipment information. The component partitioning module is used to perform variational mode decomposition on the load sequence of each power grid spatial distribution area in the spatiotemporal load data matrix, dividing it into trend sequence, seasonal sequence and residual sequence. The component model construction module is used to match external factors as input data to the trend series, seasonal series and residual series respectively. The inputs are the trend component model, seasonal component model and residual component model built based on deep learning, and the outputs are the trend forecast value, seasonal load forecast value and residual forecast mean. The fusion output module is used downstream of the trend component model, seasonal component model and residual component model to construct a physical-confidence augmented matrix containing power grid topology information using physical sensing gating units, and to generate physical gating weights using the physical-confidence augmented matrix, dynamically adjust the contribution of each component model, and output the fused total power load forecast value. The model training module is used to train each component model. After training, the parameters of each component model are fixed. The fusion model composed of each component model and the physical sensing gating unit is optimized and trained to obtain the trained fusion model. The trained fusion model is then used for power load forecasting.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-factor spatiotemporal power load forecasting method based on deep learning, characterized in that, include: S1. Collect and clean multi-source heterogeneous data, and construct a spatiotemporal load data matrix and an external factor matrix; The spatiotemporal load data matrix consists of load sequences from multiple power grid spatial distribution areas arranged by time step, and the external factor matrix consists of multiple external factors arranged by time step, including meteorological data, economic and social data, and dispatch and equipment information. S2. Perform variational mode decomposition on the load sequence of each power grid spatial distribution area in the spatiotemporal load data matrix, and divide it into trend sequence, seasonal sequence and residual sequence; S3. Match the trend series, seasonal series and residual series with external factors as input data, respectively, and input them into the trend component model, seasonal component model and residual component model built based on deep learning, respectively, and output the trend forecast value, seasonal load forecast value and residual forecast mean. S4. Downstream of the trend component model, seasonal component model and residual component model, a physical-confidence augmented matrix containing power grid topology information is constructed using a physical sensing gating unit. The physical-confidence augmented matrix is ​​then used to generate physical gating weights, dynamically adjust the contribution of each component model, and output the fused total power load forecast. S5. Train each component model, solidify the parameters of each component model after training, optimize and train the fusion model composed of each component model and the physical sensing gating unit to obtain the trained fusion model, and use the trained fusion model to predict power load.

2. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 1, characterized in that, S2 includes: Determine the initial value of the total number of modes K1 and the upper limit of the number of modes, and perform variational mode decomposition on the load sequence of each power grid spatial distribution area to construct the following constrained optimization problem: in, For the first One modal function, For the first The center frequency of each mode For Dirac functions, Denotes the square of the L2 norm. Indicates time The first-order partial derivative operator, Indicates the current moment. Represents the imaginary number symbol; Based on the current iteration round and Calculate the center frequency difference between adjacent modes ,like If it is greater than the threshold, let Proceed to the next round of optimization; if Stop iteration and output if the value is less than the threshold. As the optimal mode ; The load sequence of each power grid spatial distribution area is obtained by using variational mode decomposition. Decomposed into One mode; when and The corresponding modes are divided into trend components; the trend components are then sorted in chronological order to form a trend sequence. The sample entropy of the modality; when and Meanwhile, the period of the modal data is 24 hours or 168 hours, and the corresponding modality is divided into seasonal components; the seasonal components are sorted in chronological order to form a seasonal sequence. when and The corresponding modes are divided into residual components; the residual components are sorted in chronological order to form a residual sequence.

3. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 1, characterized in that, The construction process of the trend component model is as follows: The first input data is constructed by concatenating the trend sequence within a preset time window with long-term factors within that time window, namely GDP and weekly average temperature. The first input data and the spatial adjacency matrix are processed by a first graph convolutional network, outputting a spatial feature vector sequence. This spatial feature vector sequence is then fed into a 3-layer Transformer encoder to output the latent variables at the current time step. ; The following radial basis functions are used to calculate the similarity between latent variables: in, It is the first The implicit variables at time, express and similarity, For feature scale parameters, Denotes the square of the L2 norm. The signal variance of the trend sequence, The noise variance of the trend series; The Kronecker function; Let the feature matrix of the historical training set be... The corresponding true trend label vector is Construct the covariance matrix among historical training data based on the smoothing kernel function: in, It is the identity matrix; Build and Covariance vector between ; future Predicted mean of the time-trend component model The calculation formula is: in, It is a mean function; future Predictive dynamic variance of the time-trend component model The calculation formula is: The final trend component model outputs the future. Hourly trend forecast value and its corresponding dynamic variance .

4. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 3, characterized in that, The construction process of the seasonal component model is as follows: The second input data is constructed by concatenating periodic factors with the same frequency within a preset time window, using a 1-hour step size, for the seasonal series. These periodic factors are temperature, humidity, and irradiance. The second input data, along with the spatial adjacency matrix, is then fed into a second graph convolutional network to extract spatial features. These spatial features are then fed into a two-layer sequentially connected one-dimensional convolutional network to extract local fluctuation features, followed by a single-layer LSTM to output the future... Hourly seasonal load forecast and its corresponding static error , This represents the historical mean squared error of the seasonal component model on the validation set.

5. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 4, characterized in that, The construction process of the residual component model is as follows: Gaussian noise is injected into the residual sequence, which is then fused with random factors and fed into a third-graph convolutional network along with the spatial adjacency matrix. The random factors include sudden changes in wind speed, holiday markers, and maintenance status. The spatial feature sequence output from the third-graph convolutional network is directly fed into the GRU, and the hidden state at the GRU's terminal is mapped to a residual feature vector. ; Calculate the residual eigenvectors at different times The similarity between them is calculated using the following formula: in, It is the total number of mixed components; It is the first The mixed weights of the components; It is the first The covariance matrix of each component; It is the residual eigenvector at time i. yes and similarity, It is the first The center frequency vector of each component; Calculate the dynamic variance at the current prediction time. ,in, This is a Gram matrix constructed based on historical training data, where the element in the i-th row and j-th column is... ; yes The covariance vector between the training data and historical training data The variance of Gaussian noise is represented; future Hourly residual forecast mean The calculation formula is: ,in, It is the residual true label vector in the historical training data; Finally, the residual component model outputs the future... Hourly residual forecast mean and its corresponding dynamic variance .

6. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 5, characterized in that, The process of constructing the spatial adjacency matrix is ​​as follows: An initial adjacency matrix is ​​established based on the power grid topology. ; Calculate the Pearson correlation coefficient matrix between load sequences in different spatial distribution areas of the power grid based on historical spatiotemporal load data. Pearson correlation coefficient matrix The Line number Column elements Representative area Load time series and regional loads within the historical observation window The degree of correlation of load time series within the same historical observation window; Final spatial adjacency matrix Obtained through weighted fusion: ,in, It is a balancing factor.

7. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 5, characterized in that, The process of constructing a physical-confidence augmented matrix containing power grid topology information using a physical sensing gating unit, generating physical gating weights using the physical-confidence augmented matrix, dynamically adjusting the contribution of each component model, and outputting the fused total power load forecast includes: The physical sensing gating unit inputs a physical-confidence augmented matrix. To dynamically adjust the contribution of each component model, among which, The reactive power compensation state vector represents the real-time reactive power compensation device capacity data for each region's bus node. This is the capacity threshold vector, where represents the upper limit of the rated capacity of transformers in each region; Based on the constructed physical-confidence augmented matrix The formula for generating physical gating weights is: in, For dynamic physical weight vectors, These correspond to the gating coefficients for the trend component, seasonal component, and residual component, respectively. Both are weight matrices. Both are bias vectors. for Activation function; The combined total power load forecast is ,in, This indicates dot product.

8. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 1, characterized in that, The training of each component model includes: The Adam optimizer was used with a learning rate of 0.001 and a batch size of 64. Each component model was trained in parallel on a GPU cluster. The optimization objective was to minimize the component loss function. in, For the trend component loss, negative log-likelihood loss is used; To account for residual component loss, maximum likelihood estimation or variational inference of the objective is employed; For seasonal component loss and , These are the weighting coefficients. The first data-driven term is defined by the following formula: in, This refers to the training batch size; The time window length of the seasonal series; For the first The true seasonal load components at time t obtained from each training batch through variational mode decomposition; For the first The seasonal load forecast output at time t of the seasonal component model in each training batch; The first physical residual term is defined by the following formula: in, This represents the seasonal load forecast output by the seasonal component model at time t. This represents the total number of time steps involved in the physical loss calculation. Equivalent indoor temperature Outdoor temperature For energy efficiency ratio, For equivalent thermal resistance, For equivalent heat capacity, This represents the square of the L2 norm.

9. The multi-factor spatiotemporal power load forecasting method based on deep learning according to claim 8, characterized in that, The optimization training of the fusion model composed of each component model and the physical perception gating unit includes: After training the three component models, the parameters of all component models are fixed, and the learnable parameters of the physical perception gating unit are iteratively adjusted. The value of the overall optimization objective function is calculated during each iteration until the value of the overall optimization objective function is minimized or the preset number of iterations is reached. Then, the iteration stops, and the trained fusion model is obtained. The formula is as follows: in, and For balancing weighting coefficients; As the second data-driven term, the weighted mean square error is used to calculate the fused total power load forecast. Compared with the actual load value Deviation; The second physical residual term is defined by the following formula: in, For the number of regions, To predict the time window, For a moment Total power supply For the estimated network loss value, For the region exist The actual load value at any given time.

10. A system employing the deep learning-based multi-factor spatiotemporal power load forecasting method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect and clean multi-source heterogeneous data, and construct a spatiotemporal load data matrix and an external factor matrix; The spatiotemporal load data matrix consists of load sequences from multiple power grid spatial distribution areas arranged by time step, and the external factor matrix consists of multiple external factors arranged by time step, including meteorological data, economic and social data, and dispatch and equipment information. The component partitioning module is used to perform variational mode decomposition on the load sequence of each power grid spatial distribution area in the spatiotemporal load data matrix, dividing it into trend sequence, seasonal sequence and residual sequence. The component model construction module is used to match external factors as input data to the trend series, seasonal series and residual series respectively. The inputs are the trend component model, seasonal component model and residual component model built based on deep learning, and the outputs are the trend forecast value, seasonal load forecast value and residual forecast mean. The fusion output module is used downstream of the trend component model, seasonal component model and residual component model to construct a physical-confidence augmented matrix containing power grid topology information using physical sensing gating units, and to generate physical gating weights using the physical-confidence augmented matrix, dynamically adjust the contribution of each component model, and output the fused total power load forecast value. The model training module is used to train each component model. After training, the parameters of each component model are fixed. The fusion model composed of each component model and the physical sensing gating unit is optimized and trained to obtain the trained fusion model. The trained fusion model is then used for power load forecasting.

Images