Power Load Forecasting Method and System Based on Meteorological Prediction Model

Abstract

The present invention discloses a power load forecasting method and system based on a meteorological prediction model, which relates to the technical field of load forecasting, and includes the following steps: obtaining historical power load data and corresponding multi-dimensional meteorological data, constructing a meteorological coupling effect analysis model, and calculating a dynamic coupling weight matrix between each meteorological parameter through an attention mechanism network; performing feature fusion on the coupling weight matrix and the multi-dimensional meteorological data through a gated recurrent unit to generate a spatio-temporal coupling feature vector; establishing a power load forecasting model, and extracting short-term fluctuation features and long-term trend features of the power load from the spatio-temporal coupling feature vector based on a deep residual network architecture and parallel convolutional channels; updating the dynamic coupling weight matrix based on the prediction error backpropagation to realize the adaptive adjustment of the coupling relationship of the meteorological parameters. The present invention realizes the dynamic fusion of spatio-temporal features through a cross-gated unit, enabling the model to simultaneously utilize spatio-temporal information for power load forecasting and improving the forecasting accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and more specifically, to a power load forecasting method and system based on a meteorological forecasting model. Background Art

[0002] In today's society, the stability and reliability of power supply play a crucial supporting role in economic development and people's daily lives. With the accelerating process of global industrialization and urbanization, the scale of power demand continues to climb, and its fluctuation characteristics become increasingly complex. For the safe and efficient operation of the power system, accurate prediction of power load has become an essential link, which can provide key basis for power generation plan formulation, power grid dispatching, and infrastructure planning.

[0003] There is a close connection between meteorological conditions and power load. In terms of temperature, when the temperature deviates from the human comfort range, whether it is the large operation of air conditioning and refrigeration equipment in summer due to heat, or the wide use of heating equipment in winter due to cold, it will cause significant changes in power load. Research shows that in some high-temperature areas, for every 1°C increase in summer temperature, the power load may increase by several megawatts. Humidity cannot be ignored either. In a high-humidity environment, people may use dehumidification equipment, and humidity affects the heat dissipation of electrical equipment, indirectly changing the electricity demand. In addition, meteorological factors such as wind speed and sunshine duration also affect the electricity consumption of places relying on natural ventilation and the power load of solar photovoltaic power generation connected to the power grid.

[0004] For example, a method and system for predicting the stability of a power system based on an industrial power load simulation model disclosed in the invention patent announcement with the publication number: CN113868817A. The method includes constructing an equivalent circuit of the industrial equipment based on the voltage and phase current of the industrial equipment; using Kirchhoff's theorem to obtain an industrial power load simulation model based on the equivalent circuit of the industrial equipment; predicting the three-phase voltage, three-phase current, and power data of the industrial equipment based on the real-time three-phase current data of the industrial equipment; and evaluating whether the power system is stable according to the three-phase voltage, three-phase current, and power data of the industrial equipment.

[0005] For example, a method and system for establishing a power load prediction model based on regional characteristics and meteorology announced in the invention patent announcement with the announcement number of: CN119442049A. Through the steps: S1, establishing an influence index system based on regional characteristics and meteorological data; S2, obtaining a historical sample data set based on the influence index system; S3, constructing a machine learning model based on the random forest algorithm, and optimizing the model parameters through the historical sample data set to obtain a power load prediction model; S4, evaluating the power load prediction model; S5, predicting the power load through the power load prediction model and the influence index system. The prediction is carried out using multi-dimensional attribute variables such as regional electricity consumption building parameters, outdoor meteorological parameters, time parameters, and electricity consumption parameters at the T-t moment. At the same time, a prediction model is established through the random forest algorithm, which improves the prediction efficiency and prediction accuracy.

[0006] In the above disclosed technical solution, there are at least the following technical problems: There is a coupling effect among various meteorological factors. For example, the interaction between factors such as temperature and humidity, wind speed and air pressure will affect the power load, but the existing technology often ignores these coupling effects, resulting in inaccurate prediction results. The traditional method only simply splices or fuses spatio-temporal features through a single gate, without considering the dynamic modulation of the coupling effect of meteorological parameters. In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0007] In order to overcome the above defects of the prior art, the embodiments of the present invention provide a power load prediction method and system based on a meteorological prediction model. By constructing a meteorological coupling effect analysis model, the dynamic coupling weight matrix between each meteorological parameter is calculated through an attention mechanism network to solve the problem of not considering the dynamic modulation of the coupling effect of meteorological parameters.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A power load prediction method based on a meteorological prediction model includes the following steps: obtaining historical power load data and corresponding multi-dimensional meteorological data, where the multi-dimensional meteorological data includes four meteorological parameters: temperature, humidity, wind speed, and air pressure; constructing a meteorological coupling effect analysis model, and calculating the dynamic coupling weight matrix between each meteorological parameter through an attention mechanism network; performing feature fusion on the coupling weight matrix and the multi-dimensional meteorological data through a gated recurrent unit to generate a spatio-temporal coupling feature vector; establishing a power load prediction model, and extracting the short-term fluctuation features and long-term trend features of the power load from the spatio-temporal coupling feature vector based on the deep residual network architecture and parallel convolutional channels; updating the dynamic coupling weight matrix based on the prediction error backpropagation to realize the adaptive adjustment of the coupling relationship of meteorological parameters.

[0010] In a preferred embodiment, to construct the meteorological coupling effect analysis model, a dynamic coupling weight matrix between each meteorological parameter is calculated through an attention mechanism network, specifically: a dual-stream attention mechanism is adopted and analyzed based on time series to obtain a first attention weight matrix and a second attention weight matrix; the first attention weight matrix and the second attention weight matrix are cascaded in the feature dimension to obtain a cascaded feature matrix; based on the feature matrix and the sigmoid function mapping relationship, a dynamic weight distribution coefficient matrix is obtained; based on the dynamic weight distribution coefficient matrix and the physical constraint conditions of the meteorological parameters, a constrained weight correction function is established to obtain the final coupling weight matrix.

[0011] In a preferred embodiment, the steps for obtaining the first attention weight matrix and the second attention weight matrix are as follows: First attention weight matrix: Multidimensional meteorological data is input into the first attention stream module based on time series; based on each time step in the time series, the linear correlation coefficients between each meteorological parameter are calculated; the linear correlation coefficients are normalized through a linear transformation and the softmax function to obtain the first attention weight matrix based on the linear relationship between each meteorological parameter.

[0012] Second attention weight matrix: The second attention stream module extracts features from the input meteorological data and combines them into a non-linear interaction feature vector; the similarity between the non-linear interaction feature vector and a learnable query vector is calculated and normalized through the softmax function to obtain the second attention weight matrix.

[0013] In a preferred embodiment, to generate a spatio-temporal coupling feature vector by performing feature fusion on the coupling weight matrix and multidimensional meteorological data through a gated recurrent unit, specifically: multi-scale time information features are obtained, a meteorological station topology network is constructed according to the geographical location information of the meteorological stations, and spatial association features are generated through a dynamic graph attention network; the spatial association features are input into a feature interaction layer, and the cross-gated unit controls the fusion ratio of the spatio-temporal features to generate a spatio-temporal coupling feature vector.

[0014] In a preferred embodiment, the steps for obtaining the multi-scale time information features are as follows: The dimension of the coupling weight matrix and the multidimensional meteorological data are aligned based on time steps to obtain each time step, and the meteorological data of historical time steps are subjected to convolution operations based on dilated causal convolution and a preset convolution kernel size and dilation rate to obtain multi-scale time information features.

[0015] In a preferred embodiment, generating the spatial correlation features through the dynamic graph attention network specifically includes: obtaining features through the time dimension feature extraction layer and using them as the input of the graph attention mechanism; calculating the attention weight of each node for the graph attention mechanism; and based on the attention weights, performing weighted aggregation on the features of each node to obtain the spatial correlation features.

[0016] In a preferred embodiment, controlling the fusion ratio of the spatio-temporal features through the cross-gating unit to generate the spatio-temporal coupling feature vector specifically includes: fusing the multi-scale time information features and the spatial correlation features to obtain the fused features; performing an outer product operation on the fused features to obtain the outer product matrix, and stacking the outer product matrices of all time steps to obtain the outer product tensor; performing a tensor multiplication of the preset learning weight matrix and the outer product tensor to generate the spatio-temporal interaction gating signal; extracting features from the coupling weight matrix, performing a vectorization operation on each feature to obtain the coupling vector, stacking the coupling vectors of all time steps to obtain the coupling tensor; processing the coupling tensor through a multi-layer perceptron to generate the coupling modulation gating signal; and performing weighted splicing of the spatio-temporal interaction gating signal and the coupling modulation gating signal with the original meteorological data and the coupling weight matrix respectively to obtain the spatio-temporal coupling feature vector.

[0017] In a preferred embodiment, the parallel convolutional channels extract the short-term fluctuation features and long-term trend features of the power load from the spatio-temporal coupling feature vector specifically includes: decomposing and preprocessing the input features, and compensating the time-frequency domain features for each other to obtain the short-term fluctuation stream and the long-term trend stream; splicing and fusing the features of the short-term fluctuation stream and the long-term trend stream based on the dual-path residual unit; mapping the fused features to the power load prediction value, and compressing the feature dimension to obtain the final load prediction value.

[0018] In a preferred embodiment, updating the dynamic coupling weight matrix based on the prediction error backpropagation specifically includes: calculating the mean square error between the load prediction value and the true value; obtaining the gradient of the coupling weight matrix by taking the partial derivative of the mean square error with respect to each element of the coupling weight matrix based on the chain rule; calculating the first-order moment estimate and the second-order moment estimate of the gradient; and performing bias correction on the first-order moment estimate and the second-order moment estimate to obtain the updated coupling weight matrix.

[0019] The technical effects and advantages of the power load prediction method and system based on the meteorological prediction model of the present invention:

[0020] 1. The present invention constructs a meteorological coupling effect analysis model, calculates the dynamic coupling weight matrix among various meteorological parameters through an attention mechanism network, fully considers the interaction and coupling relationship among meteorological parameters such as temperature, humidity, wind speed, and air pressure, is more comprehensive and accurate than considering the influence of each meteorological parameter on the power load alone, can capture the comprehensive influence of meteorological factors on the power load more accurately, and thus improves the prediction accuracy.

[0021] 2. The present invention updates the coupling weight matrix through prediction error backpropagation. By calculating the mean square error between the load prediction value and the true value, taking the partial derivative of the mean square error with respect to the coupling weight matrix to obtain the gradient, and correcting the bias of the first-order moment estimation and the second-order moment estimation of the gradient to update the coupling weight matrix, the model can automatically adjust the coupling relationship among meteorological parameters according to the prediction error. As the data is continuously input and the model is continuously trained, the model can adaptively learn the dynamic change relationship between meteorological parameters and the power load, and thus maintain good prediction performance under different meteorological conditions and times. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic flow chart of the power load prediction method based on the meteorological prediction model of the present invention;

[0023] Figure 2 is a schematic structural diagram of the power load prediction system based on the meteorological prediction model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] Embodiment 1 Figure 1 provides the power load prediction method based on the meteorological prediction model of the present invention, including the following steps:

[0026] S1. Obtain historical power load data and corresponding multi-dimensional meteorological data, where the multi-dimensional meteorological data includes four meteorological parameters: temperature, humidity, wind speed, and air pressure;

[0027] There will be mutual coupling among meteorological parameters. The coupling parameter is a parameter used to describe the degree of interaction and influence among multiple variables or systems. In the context of meteorology and power load, the four meteorological factors of temperature, humidity, wind speed, and air pressure are coupling parameters, and there are mutual relationships among them and between them and the power load.

[0028] Coupling of temperature and power load:

[0029] Air temperature is an important factor affecting power load. In hot weather, the use of cooling equipment such as air conditioners increases, and the power load rises; in cold weather, the use of heating equipment also causes the power load to increase. There is usually a certain functional relationship, which can be described by establishing a load-temperature model. For example, a linear or nonlinear regression model is used to determine the approximate change in power load when the temperature changes by a certain degree. The coefficient related to this change is a coupling parameter.

[0030] Coupling of humidity and power load:

[0031] Humidity affects human comfort and the operating environment of equipment. In a high-humidity environment, people may use more dehumidification equipment or change their cooling requirements for air conditioners, thus affecting the power load. At the same time, humidity also affects the heat dissipation and other performance of some electrical equipment, and thus indirectly affects the power load. The correlation coefficient between humidity and power load can be determined through experiments or data analysis as a coupling parameter to measure the degree of coupling between them.

[0032] Coupling of wind speed and power load:

[0033] The influence of wind speed on power load is mainly reflected in heat dissipation and ventilation. For some large industrial equipment or outdoor power facilities, the magnitude of wind speed affects their heat dissipation effect, and thus affects the operating power and power consumption of the equipment. In cities, wind speed may also affect the ventilation requirements of buildings, indirectly affecting the use of equipment such as air conditioners and power load. The corresponding relationship between wind speed change and power load change can be studied to determine, for example, the proportion of power load change when the wind speed increases by a certain amount as a coupling parameter.

[0034] Coupling of air pressure and power load:

[0035] The change of air pressure may be related to the change of weather systems, and thus affect people's living and production activities, resulting in changes in power load. For example, when the air pressure is low, the weather may be stuffy, and the power load may increase due to increased cooling requirements. By analyzing the fluctuation of air pressure and power load in historical data, parameters reflecting the degree of their mutual influence, such as the covariance between air pressure change and power load change, are obtained as coupling parameters.

[0036] S2. Build a meteorological coupling effect analysis model, and calculate the coupling weight matrix between meteorological parameters through an attention mechanism network;

[0037] The establishment of the meteorological coupling effect analysis model to calculate the dynamic coupling weight matrix between meteorological parameters through an attention mechanism network is specifically as follows:

[0038] Adopt a dual-stream attention mechanism and input multi-dimensional meteorological data into the first attention stream module based on the time series;

[0039] For each time step in the time series, calculate the linear correlation coefficient between each meteorological parameter;

[0040] Normalize the linear correlation coefficient through linear transformation and the softmax function to obtain the first attention weight matrix based on the linear relationship between each meteorological parameter;

[0041] Extract features from the input meteorological data through the second attention stream module and combine them into a non-linear interaction feature vector;

[0042] Calculate the similarity between the non-linear interaction feature vector and the learnable query vector, and normalize it through the softmax function to obtain the second attention weight matrix;

[0043] Perform a concatenation operation on the first attention weight matrix output by the first attention stream and the second attention weight matrix output by the second attention stream in the feature dimension to obtain the concatenated feature matrix;

[0044] Input the feature matrix into the fully connected layer, perform linear transformation, and map the output to the interval [0,1] based on the sigmoid function to obtain the dynamic weight distribution coefficient matrix;

[0045] Obtain the physical constraint conditions of the meteorological parameters, establish a weight correction function with constraints, and obtain the final coupled weight matrix.

[0046] The coupled weight matrix includes a temperature-humidity interaction term and a wind speed-pressure gradient term;

[0047] The temperature-humidity interaction term is specifically:

[0048] For each time step, first calculate the theoretical value of the saturated water vapor pressure corresponding to the temperature data;

[0049] Then obtain the actually measured humidity data and calculate the actual saturated water vapor pressure;

[0050] Calculate the ratio of the temperature change amount to the saturated water vapor pressure change amount to obtain a value reflecting the temperature-humidity interaction relationship, which is used as the element value of the temperature-humidity interaction term in the coupled weight matrix.

[0051] The theoretical value of the saturated water vapor pressure is specifically:

[0052]

[0053] The actual saturated water vapor pressure is specifically:

[0054]

[0055] Among them, is the theoretical value of saturated water vapor pressure, is the saturated water vapor pressure under standard conditions, is the exponential function, is the latent heat of vaporization of water vapor, is the specific gas constant of water vapor, is the temperature under standard conditions, is the current temperature, is the actual saturated water vapor pressure, is the relative humidity.

[0056] The wind speed - pressure gradient term is specifically:

[0057] Obtain the wind speed value at each time step, and obtain the pressure gradient by dividing the pressure difference between adjacent meteorological stations by the distance between the stations;

[0058] Calculate the covariance between the wind speed and the pressure gradient to obtain the element value of the wind speed - pressure gradient term.

[0059] The constraint conditions include:

[0060] The temperature - humidity interaction term satisfies the Clausius - Clapeyron equation constraint, the wind speed - pressure gradient term conforms to the geostrophic wind balance equation constraint, and the weight coefficients of each coupling term maintain non - negativity constraints.

[0061] The linear correlation coefficient matrix is R. After linear transformation W1*R + b1 (where W1 is the weight matrix and b1 is the bias vector), and then through the softmax function softmax(W1*R + b1), the attention weight matrix A1 is obtained.

[0062] Adopt a two - stream attention mechanism. The first attention stream calculates the linear correlation coefficient between parameters, and the second attention stream extracts non - linear interaction features;

[0063] Fuse the outputs of the two - stream attention through the feature concatenation layer to generate dynamic weight distribution coefficients;

[0064] Introduce the physical constraint conditions of meteorological parameters, establish a weight correction function with constraints, and ensure that the coupling relationship conforms to the laws of atmospheric physics

[0065] The calculation of the linear correlation coefficient between each meteorological parameter is specifically:

[0066]

[0067] Among them, is the linear correlation coefficient, The value of meteorological parameter X at time step i is the mean value of meteorological parameter X The value of meteorological parameter Y at time step i is the mean value of meteorological parameter Y is the number of time steps

[0068] S3. Feature fusion of the coupling weight matrix and multi-dimensional meteorological data is performed through a gated recurrent unit to generate a spatio-temporal coupling feature vector

[0069] The feature fusion of the coupling weight matrix and multi-dimensional meteorological data through a gated recurrent unit to generate a spatio-temporal coupling feature vector is specifically as follows

[0070] Align the dimensions of the coupling weight matrix and multi-dimensional meteorological data based on the time step

[0071] Input the original meteorological data into the time dimension feature extraction layer, and the time dimension feature extraction layer uses dilated causal convolution

[0072] For each time step, the dilated causal convolution performs a convolution operation on the meteorological data of historical time steps according to the preset convolution kernel size and dilation rate to obtain multi-scale time information features

[0073] Construct a meteorological station topological graph network according to the geographical location information of meteorological stations, and generate spatial correlation features through a dynamic graph attention network

[0074] Input the spatial correlation features into the feature interaction layer, and realize the dynamic fusion of spatio-temporal features through a cross-gated unit

[0075] The cross-gated unit controls the fusion ratio of spatio-temporal features by calculating the gating signal between time features and spatial features, so as to generate a spatio-temporal coupling feature vector

[0076] The generation of spatial correlation features through a dynamic graph attention network is specifically as follows

[0077] Use the feature representation obtained by the time dimension feature extraction layer as the input of the graph attention mechanism

[0078] For each node (meteorological station), the graph attention mechanism calculates the attention weight of the node according to the features of neighboring nodes and the weights of edges

[0079] Through the graph attention mechanism, the features of each node are weighted and aggregated to obtain spatial correlation features

[0080] The cross-gated unit controls the fusion ratio of spatio-temporal features by calculating the gating signal between time features and spatial features, so as to generate a spatio-temporal coupling feature vector, specifically as follows

[0081] Fuse the multi-scale time information features and spatial correlation features to obtain fused features;

[0082] Perform an outer product operation on the fused features to obtain an outer product matrix, and stack the outer product matrices of all time steps to obtain an outer product tensor;

[0083] Perform a tensor multiplication of a preset learning weight matrix and the outer product tensor to generate a spatio-temporal interaction gating signal;

[0084] Extract features from the coupling weight matrix, perform a vectorization operation on each feature to obtain coupling vectors, and stack the coupling vectors of all time steps to obtain a coupling tensor;

[0085] Process the coupling tensor through a multi-layer perceptron to generate a coupling modulation gating signal;

[0086] Perform weighted splicing of the spatio-temporal interaction gating signal and the coupling modulation gating signal with the original meteorological data and the coupling weight matrix respectively to obtain a spatio-temporal coupling feature vector.

[0087] In the meteorological station topology network, each node represents a meteorological station, the edges between the nodes represent the spatial correlation relationships between the stations, and the weights of the edges can be set according to factors such as the distance and terrain between the stations;

[0088] For missing values, according to the time correlation of the data, linear interpolation or predicted values based on a time series model (such as the ARIMA model) can be used for filling; for outliers, statistical methods (such as the 3σ principle) can be used for identification and replacement with reasonable values. Then, perform normalization processing on the data. For example, use the min-max normalization method to map the values of each meteorological parameter to the interval [0, 1] to ensure that meteorological parameters with different dimensions can participate in subsequent calculations on the same scale.

[0089] S4. Establish a power load prediction model, and extract the short-term fluctuation features and long-term trend features of the power load from the spatio-temporal coupling feature vector based on the deep residual network architecture and parallel convolutional channels;

[0090] The parallel convolutional channels extract the short-term fluctuation features and long-term trend features of the power load from the spatio-temporal coupling feature vector, specifically:

[0091] Decompose the input features into a high-frequency short-term fluctuation stream and a low-frequency long-term trend stream, and perform preprocessing using wavelet packet transform and moving average filtering respectively;

[0092] Perform time-frequency domain feature mutual compensation through inverse wavelet transform and fast Fourier transform, that is, use frequency domain information to enhance time domain features;

[0093] Based on dual-path residual units, the short-term fluctuation flow and the long-term trend flow are processed separately, and a cascaded dilated convolution and transposed convolution structure is adopted for each path;

[0094] The features of the short-term fluctuation flow and the long-term trend flow after being processed by the dual-path residual units are concatenated and fused;

[0095] The output layer maps the fused features to the power load prediction value, and uses a fully connected layer to compress the feature dimension to 1 to obtain the final load prediction value.

[0096] Cascaded dilated convolution: In each path, cascaded dilated convolution layers are used to capture features at different scales. Dilated convolution can increase the receptive field of the convolution kernel without increasing the number of parameters, so as to better capture long-term dependencies and short-term fluctuations. For example, different dilation rates (such as 1, 2, 4, etc.) are set to construct multiple dilated convolution layers.

[0097] Transposed convolution structure: At the end of each path, a transposed convolution layer is used to restore the dimension of the features to match the requirements of the output layer. Transposed convolution can map low-dimensional features to high-dimensional space, which helps to improve the accuracy of prediction.

[0098] S5, update the dynamic coupling weight matrix based on the backpropagation of the prediction error to achieve the adaptive adjustment of the coupling relationship of meteorological parameters.

[0099] The update of the dynamic coupling weight matrix based on the backpropagation of the prediction error is specifically as follows:

[0100] Calculate the mean square error between the load prediction value and the true value;

[0101] Based on the chain rule, take the partial derivative of the mean square error with respect to each element of the coupling weight matrix to obtain the gradient of the coupling weight matrix;

[0102] Calculate the first moment estimate (mean) and the second moment estimate (uncentered variance) of the gradient;

[0103] Perform bias correction on the first moment estimate and the second moment estimate to obtain the updated coupling weight matrix.

[0104] The calculation of the first moment estimate and the second moment estimate of the gradient is specifically as follows:

[0105]

[0106] Among them, is the first moment estimate of the gradient, is the second moment estimate of the gradient, and are the decay rates, which are 0.9 and 0.99 respectively, is the gradient of the coupling weight matrix, and are the first - order moment estimate and the second - order moment estimate of the gradient of the previous iteration respectively, is the number of iterations.

[0107] Performing bias correction on the first - order moment estimate and the second - order moment estimate to obtain an updated coupled weight matrix, specifically:

[0108]

[0109]

[0110] wherein, is the first - order moment estimate after bias correction, is the second - order moment estimate after bias correction, and are bias correction factors respectively, is the updated coupled weight matrix, is the coupled weight matrix before update, is a very small constant, is the learning rate, is the number of iterations.

[0111] Example 2, A power load forecasting system based on a meteorological forecasting model, includes the following modules:

[0112] Data acquisition module: Used to acquire historical power load data and corresponding multi - dimensional meteorological data, and the multi - dimensional meteorological data includes four meteorological parameters: temperature, humidity, wind speed, and air pressure;

[0113] Meteorological coupling weight calculation module: Used to construct a meteorological coupling effect analysis model and calculate the dynamic coupling weight matrix between each meteorological parameter through an attention mechanism network;

[0114] Spatio - temporal feature fusion module: Used to perform feature fusion on the coupling weight matrix and the multi - dimensional meteorological data through a gated recurrent unit to generate a spatio - temporal coupling feature vector;

[0115] Power load forecasting and modeling module: Used to establish a power load forecasting model, and extract the short - term fluctuation features and long - term trend features of the power load from the spatio - temporal coupling feature vector based on a deep residual network architecture and parallel convolutional channels;

[0116] Coupled weight adaptive update module: Used to update the dynamic coupling weight matrix based on the prediction error backpropagation to realize the adaptive adjustment of the coupling relationship of meteorological parameters.

[0117] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0118] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product.

[0119] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0120] In addition, the functional modules in each embodiment of this application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0121] As mentioned above, the above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application, and all should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0122] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting power load based on a meteorological prediction model, characterized in that: The following steps are involved: Acquire historical power load data and corresponding multi-dimensional meteorological data, wherein the multi-dimensional meteorological data includes four meteorological parameters: temperature, humidity, wind speed, and air pressure; Construct a meteorological coupling effect analysis model and calculate the dynamic coupling weight matrix between meteorological parameters through the attention mechanism network; The calculation of the dynamic coupling weight matrix between the meteorological parameters is specifically as follows: A dual-stream attention mechanism is used, and the first attention weight matrix and the second attention weight matrix are obtained based on time series analysis; Perform a cascade operation on the first attention weight matrix and the second attention weight matrix in the feature dimension to obtain a cascaded feature matrix; Based on the mapping relationship between the feature matrix and the sigmoid function, the dynamic weight allocation coefficient matrix is ​​obtained; Based on the dynamic weight allocation coefficient matrix and the physical constraints of meteorological parameters, a constrained weight correction function is established to obtain the final coupling weight matrix. The coupling weight matrix and multi-dimensional meteorological data are fused through gated recurrent units to generate spatiotemporal coupling feature vectors. Establish a power load forecasting model to extract short-term fluctuation characteristics and long-term trend characteristics of power load from spatiotemporal coupling feature vectors based on a deep residual network architecture and parallel convolution channels; The generating of the spatiotemporal coupling characteristic vector is specifically as follows: Obtain multi-scale temporal information features, build a meteorological station topology network based on the geographical location information of the meteorological station, and generate spatial correlation features through a dynamic graph attention network; The spatial correlation features are input into the feature interaction layer, and the multi-scale time information features and the spatial correlation features are fused to obtain fused features; Perform outer product operation on the fused features to obtain the outer product matrix, and stack the outer product matrices of all time steps to obtain the outer product tensor; Perform tensor multiplication of the preset learning weight matrix and the outer product tensor to generate a spatiotemporal interactive gating signal; Extract features from the coupling weight matrix, perform vectorization operations on each feature to obtain a coupling vector, and stack the coupling vectors of all time steps to obtain a coupling tensor; The coupling tensor is processed by a multi-layer perceptron to generate a coupling modulation gating signal; The spatiotemporal interaction gating signal and the coupled modulation gating signal are weightedly concatenated with the original meteorological data and the coupling weight matrix to obtain the spatiotemporal coupling feature vector. The dynamic coupling weight matrix is ​​updated based on the back propagation of prediction errors to achieve adaptive adjustment of the coupling relationship of meteorological parameters.

2. The method for predicting power load based on a meteorological prediction model according to claim 1, characterized in that: The specific steps of obtaining the first attention weight matrix and the second attention weight matrix are as follows: First attention weight matrix: Input the multidimensional meteorological data into the first attention flow module based on the time series; Based on each time step in the time series, the linear correlation coefficient between each meteorological parameter is calculated; The linear correlation coefficient is normalized by linear transformation and softmax function to obtain a first attention weight matrix based on the linear relationship between the meteorological parameters; Second attention weight matrix: The input meteorological data is feature extracted through the second attention flow module and combined into a nonlinear interaction feature vector; The second attention weight matrix is ​​obtained by calculating the similarity between the nonlinear interaction feature vector and the learnable query vector and normalizing it with the softmax function.

3. The method for predicting power load based on a meteorological prediction model according to claim 2, characterized in that: The specific steps of obtaining multi-scale time information features are as follows: The coupling weight matrix dimensions and multidimensional meteorological data are aligned based on the time step to obtain each time step, and based on the dilated causal convolution and the preset convolution kernel size and dilation rate, the meteorological data of the historical time step is convolved to obtain the multi-scale time information features.

4. The method for predicting power load based on a meteorological prediction model according to claim 3, characterized in that: The spatial correlation features are generated by the dynamic graph attention network, specifically: Features are obtained through the time dimension feature extraction layer and used as the input of the graph attention mechanism; For each node and graph attention mechanism, calculate the attention weight of the node; Based on the attention weight, the features of each node are weighted and aggregated to obtain spatial correlation features.

5. The method for predicting power load based on a meteorological prediction model according to claim 4, characterized in that: The parallel convolution channel extracts the short-term fluctuation characteristics and long-term trend characteristics of the power load from the spatiotemporal coupling feature vector, specifically: Decompose and preprocess the input features, and compensate the time-frequency domain features to obtain short-term fluctuation flow and long-term trend flow; Based on the dual-path residual unit, the features of short-term fluctuation flow and long-term trend flow are spliced ​​and fused; The fused features are mapped to the power load forecast value, and the feature dimensions are compressed to obtain the final load forecast value.

6. The method for predicting power load based on a meteorological prediction model according to claim 5, characterized in that: The dynamic coupling weight matrix is ​​updated based on the prediction error back propagation, specifically: Calculate the mean square error between the load forecast value and the actual value; Based on the chain rule, the partial derivative of the mean square error with respect to each element of the coupling weight matrix is ​​calculated to obtain the gradient of the coupling weight matrix; Compute first-order and second-order moment estimates of the gradient; The first-order moment estimation and the second-order moment estimation are bias-corrected to obtain an updated coupling weight matrix.

7. A system for the method for predicting power load based on a meteorological prediction model according to any one of claims 1 to 6, characterized in that: Includes the following modules: Data acquisition module: used to obtain historical power load data and corresponding multi-dimensional meteorological data, the multi-dimensional meteorological data including four meteorological parameters: temperature, humidity, wind speed and air pressure; Meteorological coupling weight calculation module: used to build a meteorological coupling effect analysis model and calculate the dynamic coupling weight matrix between meteorological parameters through the attention mechanism network; Spatiotemporal feature fusion module: used to fuse the coupling weight matrix and multidimensional meteorological data through a gated recurrent unit to generate a spatiotemporal coupling feature vector; Power load forecasting modeling module: used to establish a power load forecasting model, extracting short-term fluctuation characteristics and long-term trend characteristics of power load from spatiotemporal coupling feature vectors based on a deep residual network architecture and parallel convolution channels; Coupling weight adaptive update module: used to update the dynamic coupling weight matrix based on the back propagation of prediction errors to achieve adaptive adjustment of the coupling relationship of meteorological parameters.

Images