In-province power load prediction method based on multi-modal information fusion analysis

Abstract

The invention relates to the technical field of electric quantity prediction, and discloses an in-province power load prediction method based on multi-modal information fusion analysis, and the method comprises the steps: carrying out the preprocessing of collected historical load data and market data, extracting the Seaonal seasonal features and trend trend features of the data through a time sequence information extraction module, and carrying out the prediction of the Seaonal seasonal features and the trend trend features of the data; utilizing a TextCNN model to extract public opinion features of the market public opinion corpus; then, multi-angle analysis is performed on the three types of features through a multi-head attention module, a combined feature vector is obtained through normalization, affine transformation, dynamic weight calculation and weighted feature fusion, after the combined feature vector is input into a gate control unit for processing, a prediction result is output through a feedforward neural network, and finally, the prediction accuracy is measured by adopting MAE and RMSE indexes. According to the method, multi-dimensional features are fused, a mode of combining deep learning and machine learning is adopted, the precision, stability and anti-interference capability of power load prediction are effectively improved, and reliable support is provided for operation and management of a power system.

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting technology, specifically a method for forecasting provincial power load based on multimodal information fusion analysis. Background Technology

[0002] In the complex architecture of modern power systems, short-term load forecasting plays an irreplaceable role as a crucial link in ensuring the stable and efficient operation of the system. Short-term load forecasting is typically measured in hourly or daily units, and its significance is profound. It not only provides core data for power dispatching departments to rationally arrange power generation plans and optimize grid operation, helping to achieve precise matching of power supply and demand and reduce power generation costs and grid losses; but also, in the context of the power market, it provides strong support for the strategy formulation of power trading participants, enhancing the transparency and stability of market transactions.

[0003] Current power load forecasting methods primarily involve collecting and selecting historical load data, constructing load forecasting models based on statistical data, and ensuring forecast accuracy through effective model selection and construction. However, these existing methods have some significant limitations: they lack consideration of the impact of the interaction of various market information data on power load, and ignore recent policy regulations, announcements, major economic activities, and other market sentiment data. Furthermore, power data typically exhibits obvious seasonality, periodicity, and trend characteristics, and existing methods fail to fully utilize the spatiotemporal correlations within power data, resulting in room for improvement in forecast accuracy and stability. Therefore, to improve the accuracy of power load forecasting across the province and better meet the operation and management needs of the power system, a new power load forecasting method that combines textual data with complex, correlated market data is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to provide a provincial power load forecasting method based on multimodal information fusion analysis, which aims to improve the accuracy, stability and anti-interference capability of power load forecasting by fusing multi-dimensional features and combining deep learning and machine learning.

[0005] Specifically, the following steps are included:

[0006] A method for forecasting provincial power load based on multimodal information fusion analysis includes:

[0007] S1. Data Collection and Preprocessing: Collect historical load data, market data, and market sentiment data. Apply max-min normalization to the historical load data. The preprocessing formula is:

[0008]

[0009] in, For the i-th data after normalization, For the original data set, The minimum value of the data set. The maximum value in the dataset; the market data includes hourly meteorological data such as temperature, humidity, wind speed, and sunshine, as well as urban economic data, power transmission line data, and holiday schedule information; during data preprocessing;

[0010] During data preprocessing, date information is encoded into discrete values, text news information and industry information are collected, the corpus is segmented and word vectorized, and a dataset containing multiple features is constructed.

[0011] S2. Model training data partitioning: The preprocessed dataset is divided into training and validation sets in an 8:2 ratio. The mean squared error (MSE) function is selected as the loss function for model training. The error is reduced by adjusting the model parameters. The model parameters are saved when the error in the validation set is the lowest.

[0012] S3. Deep Learning Model Construction and Feature Extraction: The deep learning model includes a temporal information extraction module and a multi-head attention module. The temporal information extraction module is used to extract seasonal features and trend features.

[0013] S3.1. Seasonal feature extraction: Local features of the load are calculated by convolution with kernel dimension [1,T] and stride of 1, and global features of the load are calculated by convolution with kernel dimension [1,T] and sliding window stride of T.

[0014] A two-dimensional convolution operation with a kernel dimension of [1, feature_size] and a sliding window stride of 1 is adopted, where feature_size represents the total number of feature dimensions in the input data, which is used to dynamically assign weights to different market information features, thereby suppressing the influence of unimportant features;

[0015] Aligning and stitching together the local and global features at different scales allows the model to focus more effectively on key feature information.

[0016] The formula for calculating one-dimensional convolution is:

[0017] The i-th element of the convolution result. Let be the element at time point t of the input sequence. Let be the it-th element of the convolution kernel, and T be the sequence period;

[0018] The formula for calculating two-dimensional convolution is:

[0019]

[0020] in, For elements of the two-dimensional convolution result, For input values, Elements of a two-dimensional convolution kernel;

[0021] Merge and align convolution results:

[0022] in, These are the calculation results of different convolutions;

[0023] S3.2.trend Feature Extraction: Removes non-stationary information from the input sequence, especially the mean and standard deviation of instances, and returns the information removed by input normalization to the model to achieve the model's inverse normalization operation.

[0024] Specifically, the sequence is decomposed into the following parts through the time-series information extraction module: In the formula, For the observation at time t, Represents the trend component at time t; This represents the seasonal component at time t; This represents the residual term at time t;

[0025] Trend Components The calculation expression is:

[0026]

[0027] Seasonal components The calculation expression is: In the formula, , , These represent the y values ​​at times t-1, t-v+1, and tv, respectively. Represents a data set The observation value at the i-th time point, Represents a data set The average of all observations; This represents the average value across all points in time within the dataset. Indicates the length of seasonal cycles; These represent the trend constant and the trend slope, respectively. This represents the time point corresponding to the i-th observation.

[0028] S4. Public opinion feature extraction: Market public opinion corpus information is converted into word vectors through an embedding layer. After training the TextCNN model, the encoded results are input into a fully connected layer, and the public opinion features are output through the sigmoid function.

[0029] S5. Multi-head Attention Calculation: The multi-head attention module performs multi-head attention calculation on the trend features and seasonal features extracted in S3 and the public opinion features extracted in S4, calculating the query, key, and value matrix.

[0030]

[0031]

[0032] Among them, among them, , and This is the weight matrix; further, the attention score is calculated and expressed as:

[0033]

[0034] in, It is the dimension of the query and key vector; through Calculate multi-head attention, where It is a weight matrix. It is the output projection matrix; passed through the feedforward neural network layer Processing, among which, Weight matrix, , It is the bias vector;

[0035] S6. Feature Fusion and Prediction: Normalize and perform affine transformation on the three types of output features, and calculate the dynamic weights between each feature.

[0036] in,

[0037]

[0038]

[0039] , These represent seasonal features, public opinion features, and trend features, respectively. It is a learnable projection matrix;

[0040] These are the mean and standard deviation of the features. These are learnable scaling and offset parameters; through Perform weighted feature fusion, where t represents the time step. This represents the fused feature vector; the fused features are input to the gate control unit for processing, and the gate control is calculated as follows:

[0041]

[0042] The state update formula is The gate control output is calculated as follows:

[0043] in, , , These represent the input gate, forget gate, output gate, and candidate state, respectively. denoted by sigmoid, and tanh by hyperbolic tangent activation function. This represents the state vector from the previous time step. This represents the bias vector; finally, the prediction result is output through a feedforward neural network. ;

[0044] S7. Prediction Accuracy Assessment: Prediction accuracy is measured using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics, calculated using the following formulas:

[0045]

[0046]

[0047] in, and Let i be the actual value and the predicted value at time i, respectively, and n be the number of data points.

[0048] Preferably, the market data mentioned in S1 includes hourly meteorological data such as temperature, humidity, wind speed, and sunshine, as well as urban economic data, power transmission line data, and holiday arrangement information; during data preprocessing, date information is encoded as categorical variables to construct a dataset containing multiple features.

[0049] Preferably, the calculation formula for the one-dimensional convolution operation in S3.1 is as follows: ,in, The i-th element of the convolution result. Let be the element at time point t of the input sequence. Let be the ith element of the convolution kernel, and T be the sequence period; the formula for calculating the two-dimensional convolution operation is: ,in, For elements of the two-dimensional convolution result, For input values, The elements of the 2D convolution kernel are used; the result of merging and aligning the convolutions is... , , These are the calculation results of different convolutions.

[0050] Preferably, the public opinion corpus information in S4 is collected through multiple channels and converted into word vectors after preprocessing. The word vectors contain information about positive, negative, or neutral market attitudes.

[0051] Preferably, during model training, the number of hidden layer neurons is set to 64, the learning rate is 0.001, and the Adam optimizer is used for parameter updates.

[0052] The beneficial effects of this invention are as follows:

[0053] 1. Based on market information and historical power load values, this invention obtains seasonal features, trend features, and public opinion features by performing time-series decomposition of external market information and public opinion characteristics. It then uses a combination of deep learning and machine learning to fuse and derive prediction results, which can effectively improve overall prediction performance, enhance the model's anti-interference ability, and improve prediction accuracy.

[0054] 2. By combining market sentiment information with complex market data, this invention enhances the accuracy and robustness of load forecasting, effectively captures the impact of market information fluctuations on power load, and improves forecast stability. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the provincial power load forecasting method based on multimodal information fusion, according to an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram of the multimodal information fusion process according to an embodiment of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0059] S1. Data Collection and Preprocessing: Collect historical load data, market data, and market sentiment data. Apply max-min normalization to the historical load data. The preprocessing formula is:

[0060]

[0061] in, For the i-th data after normalization, For the original data set, The minimum value of the data set. The maximum value of the dataset; the market data includes hourly temperature, humidity, wind speed, sunshine and other meteorological data, as well as urban economic data, power transmission line data and holiday arrangement information; during data preprocessing, date information is encoded into discrete values, text news information and industry information are collected, the corpus is segmented and word vectorized, and a dataset containing multiple features is constructed.

[0062] S2. Model training data partitioning: The preprocessed dataset is divided into training and validation sets in an 8:2 ratio. The mean squared error (MSE) function is selected as the loss function for model training. The error is reduced by adjusting the model parameters. The model parameters are saved when the error in the validation set is the lowest.

[0063] S3. Deep Learning Model Construction and Feature Extraction: The deep learning model includes a temporal information extraction module and a multi-head attention module. The temporal information extraction module is used to extract seasonal features and trend features.

[0064] S3.1. Seasonal Feature Extraction: Local features are calculated using convolutions with kernel dimensions [1,T] and stride of 1, while global features are calculated using convolutions with kernel dimensions [1,T] and sliding window stride of T. A two-dimensional convolution operation with kernel dimensions [1,feature_size] and sliding window stride of 1 is used, where feature_size represents the total number of feature dimensions in the input data. This is used to dynamically assign weights to different market information features, thereby suppressing the influence of unimportant features. The aforementioned local and global features at different scales are aligned and concatenated, enabling the model to more effectively focus on key feature information.

[0065] The formula for calculating one-dimensional convolution is:

[0066] The i-th element of the convolution result. Let be the element at time point t of the input sequence. Let be the it-th element of the convolution kernel, and T be the sequence period;

[0067] The formula for calculating two-dimensional convolution is:

[0068]

[0069] in, For elements of the two-dimensional convolution result, For input values, Elements of a two-dimensional convolution kernel;

[0070] Merge and align convolution results:

[0071] in, These are the calculation results of different convolutions;

[0072] S3.2.trend Feature Extraction: Removes non-stationary information from the input sequence, especially the mean and standard deviation of instances, and returns the information removed by input normalization to the model to achieve the model's inverse normalization operation.

[0073] Specifically, the sequence is decomposed into the following parts through the time-series information extraction module: In the formula, For the observation at time t, Represents the trend component at time t; This represents the seasonal component at time t; This represents the residual term at time t;

[0074] Trend Components The calculation expression is:

[0075]

[0076] Seasonal components The calculation expression is: In the formula, , , These represent the y values ​​at times t-1, t-v+1, and tv, respectively. Represents a data set The observation value at the i-th time point, Represents a data set The average of all observations; This represents the average value across all points in time within the dataset. Indicates the length of seasonal cycles; These represent the trend constant and the trend slope, respectively. This represents the time point corresponding to the i-th observation.

[0077] S4. Public opinion feature extraction: Market public opinion corpus information is converted into word vectors through an embedding layer;

[0078] After training with the TextCNN model, the encoded results are input into a fully connected layer, and the sentiment features are output through the sigmoid function.

[0079] S5. Multi-head Attention Calculation: The multi-head attention module performs multi-head attention calculation on the trend features and seasonal features extracted in S3 and the public opinion features extracted in S4, calculating the query, key, and value matrix.

[0080]

[0081]

[0082] Among them, among them, , and This is the weight matrix; further, the attention score is calculated and expressed as:

[0083]

[0084] in, It is the dimension of the query and key vector; through Calculate multi-head attention, where It is a weight matrix. It is the output projection matrix; passed through the feedforward neural network layer Processing, among which, It is a weight matrix. , It is the bias vector;

[0085] S6. Feature Fusion and Prediction: Normalize and perform affine transformation on the three types of output features, and calculate the dynamic weights between each feature.

[0086] in,

[0087]

[0088]

[0089] , These represent seasonal features, public opinion features, and trend features, respectively. It is a learnable projection matrix;

[0090] These are the mean and standard deviation of the features. These are learnable scaling and offset parameters; through Perform weighted feature fusion, where t represents the time step. This represents the fused feature vector; the fused features are input to the gate control unit for processing, and the gate control is calculated as follows:

[0091]

[0092] The state update formula is The gate control output is calculated as follows: ;

[0093] in, , , These represent the input gate, forget gate, output gate, and candidate state, respectively. denoted by sigmoid, and tanh by hyperbolic tangent activation function. This represents the state vector from the previous time step. This represents the bias vector; finally, the prediction result is output through a feedforward neural network. ;

[0094] S7. Prediction Accuracy Assessment: Prediction accuracy is measured using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics, calculated using the following formulas:

[0095]

[0096]

[0097] in, and Let i be the actual value and the predicted value at time i, respectively, and n be the number of data points.

[0098] The following also needs to be explained based on the above embodiments:

[0099] Historical load data are processed using max-min normalization. The preprocessing formula is as follows:

[0100]

[0101] in, For the i-th data after normalization, For the original data set, The minimum value of the data set. The maximum value in the dataset;

[0102] Data collection scope: Historical load data (sampling granularity of 15 minutes, covering all time periods of weekdays, weekends, and holidays); two types of market data are also collected simultaneously: first, hourly meteorological sensing data such as temperature, humidity, wind speed, sunshine duration, and rainfall from various monitoring stations within the province; second, socio-economic and power grid operation data such as the province's monthly GDP growth rate, electricity application data for key industrial parks, transmission line loss rate, main transformer load rate, and statutory holidays / workday adjustments; in addition, market sentiment data are also collected from the same period through channels such as the power trading platform bulletin board, local government websites, and industry news media.

[0103] Preprocessing operations: Max-min normalization is performed on historical load data to eliminate dimensional differences; linear interpolation is used to complete missing meteorological data; load anomalies caused by faults are identified and removed using the 3σ principle; date information is encoded as categorical variables, such as Monday to Sunday being encoded as 1-7, and statutory holidays being marked as special identifiers, and finally a standardized dataset containing 23 basic features is constructed.

[0104] The preprocessed dataset was randomly divided into training and validation sets in an 8:2 ratio, ensuring that the temporal distribution characteristics of the two datasets were consistent, covering all four seasons and various electricity consumption scenarios. Mean squared error was chosen as the loss function for model training, and an early stopping strategy was adopted during training: if the validation set error did not decrease for 15 consecutive training cycles, training was stopped, and the model parameters at this point were saved to avoid overfitting.

[0105] The core of the model includes a time-series information extraction module and a multi-head attention module. The time-series information extraction module is responsible for extracting the seasonal and trend features of the load data.

[0106] Seasonal feature extraction: A multi-scale convolution strategy is adopted to capture features in layers. First, local fluctuation features of intraday load and global seasonal patterns across cycles are extracted using convolution of specific specifications. Then, dynamic weights are assigned to different market features through two-dimensional convolution to reduce the weight ratio of weakly correlated features such as rainfall. Finally, the local and global load features are spliced ​​and aligned with the weighted market features to allow the model to focus on the core seasonal patterns.

[0107] Trend feature extraction: First, the load sequence is decomposed into trend components, seasonal components and random residual terms through the time series information extraction module. The trend-related parameters are solved by linear fitting. Then, the mean, standard deviation and other information removed during the normalization process are fed back to the model to realize the dimensional restoration of the trend features and ensure the accuracy of subsequent predictions.

[0108] Corpus preprocessing: The collected public opinion corpus is deduplicated, segmented, and stop words are removed. Then, the text is converted into 128-dimensional word vectors through the Word2Vec embedding layer. These word vectors can represent the positive, negative, or neutral attitude of the market. For example, "new energy project implementation" corresponds to a positive attitude, and "production restriction of high energy-consuming enterprises" corresponds to a negative attitude.

[0109] Feature quantization: The word vectors are input into the TextCNN model, and three different sizes of convolutional kernels are used to extract local semantic features of the text. After dimensionality reduction by max pooling, the encoded results are sent to a fully connected layer, and finally a 32-dimensional public opinion feature vector is output through the sigmoid function to complete the quantitative representation of public opinion information.

[0110] Feature matrix mapping: The vectors formed by concatenating seasonal features, trend features, and public opinion features are mapped into three types of matrices: query, key, and value, respectively, to achieve dimensional transformation and preliminary processing of features.

[0111] Attention weight calculation: The correlation weights between the three types of features are calculated by scaling the dot product to alleviate the gradient vanishing problem caused by the high feature dimensionality; 8 attention heads are set to fuse features from multiple perspectives, and then the fused features are further transformed by a feedforward neural network.

[0112] Feature normalization and weight allocation: First, layer normalization and affine transformation are performed on the three types of features. Then, the dynamic weights of each type of feature are calculated through a multi-head attention mechanism to achieve differentiated distinction of the importance of different features.

[0113] Weighted fusion and gating: Based on dynamic weights, the three types of features are weighted and fused to obtain the combined feature vector at each time step; the fused features are input into the gating control unit, and selective memory and transmission of features are achieved through input gate, forget gate, output gate and candidate state to complete the deep processing of features.

[0114] Prediction output: The features processed by the gate control unit are fed into the feedforward neural network, and the final output is the prediction result of the power load.

[0115] The prediction accuracy is measured using both mean absolute error and root mean square error. Actual test data shows that the model can control the mean absolute error of daily load prediction within the province to within 12MW and the root mean square error to below 18MW. Compared with traditional prediction models, the two error indicators are reduced by 32% and 28% respectively, which can meet the accuracy requirements of power dispatching.

[0116] The market data mentioned in S1 includes hourly meteorological data such as temperature, humidity, wind speed, and sunshine, as well as urban economic data, power transmission line data, and holiday schedule information; during data preprocessing, date information is encoded as categorical variables to construct a dataset containing multiple features.

[0117] The formula for calculating one-dimensional convolution in S3.1 is as follows: ,in, The i-th element of the convolution result. Let be the element at time point t of the input sequence. Let be the ith element of the convolution kernel, and T be the sequence period; the formula for calculating the two-dimensional convolution operation is: ,in, For elements of the two-dimensional convolution result, For input values, The elements of the 2D convolution kernel are used; the result of merging and aligning the convolutions is... , , These are the calculation results of different convolutions.

[0118] The public opinion corpus information in S4 is collected through multiple channels and converted into word vectors after preprocessing. The word vectors contain information on positive, negative or neutral market attitudes.

[0119] During model training, the number of hidden layer neurons was set to 64, the learning rate was 0.001, and the Adam optimizer was used for parameter updates.

[0120] The following also needs to be explained based on the above embodiments:

[0121] 1. Hyperparameter configuration: The number of hidden layer neurons in the model training is set to 64, the learning rate is 0.001, the batch size is 32, the Adam optimizer is used for parameter updates, and the maximum number of training iterations is 100 epochs.

[0122] 2. Timeliness of public opinion processing: A 7-day sliding time window is set to include public opinion information within 7 days before the prediction time in the feature extraction scope, ensuring the time correlation between public opinion features and load changes.

[0123] 3. Model Deployment and Application: Establish a real-time data access interface to realize the automatic collection and preprocessing of historical load, meteorological data, and public opinion information. The prediction results can be pushed to the power dispatching system at the hourly or daily level, providing support for power generation planning and power grid risk early warning.

[0124] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting provincial power load based on multimodal information fusion analysis, characterized in that, include: S1. Data Collection and Preprocessing: Collect historical load data, market data, and market sentiment data. Apply max-min normalization to the historical load data. The preprocessing formula is: in, For the i-th data after normalization, For the original data set, The minimum value of the data set. The maximum value in the dataset; the market data includes hourly meteorological data such as temperature, humidity, wind speed, and sunshine, as well as urban economic data, power transmission line data, and holiday schedule information; During data preprocessing, date information is encoded into discrete values, text news information and industry information are collected, the corpus is segmented and word vectorized, and a dataset containing multiple features is constructed. S2. Model training data partitioning: The preprocessed dataset is divided into training and validation sets in an 8:2 ratio. The mean squared error (MSE) function is selected as the loss function for model training. The error is reduced by adjusting the model parameters. The model parameters are saved when the error in the validation set is the lowest. S3. Deep Learning Model Construction and Feature Extraction: The deep learning model includes a temporal information extraction module and a multi-head attention module. The temporal information extraction module is used to extract seasonal features and trend features. S3.

1. Seasonal feature extraction: Local features of the load are calculated by convolution with kernel dimension [1,T] and stride of 1, and global features of the load are calculated by convolution with kernel dimension [1,T] and sliding window stride of T. A two-dimensional convolution operation with a kernel dimension of [1, feature_size] and a sliding window stride of 1 is adopted, where feature_size represents the total number of feature dimensions in the input data, which is used to dynamically assign weights to different market information features, thereby suppressing the influence of unimportant features; Aligning and stitching together the local and global features at different scales allows the model to focus more effectively on key feature information. The formula for calculating one-dimensional convolution is: The i-th element of the convolution result Let be the element at time point t of the input sequence. Let be the it-th element of the convolution kernel, and T be the sequence period; The formula for calculating two-dimensional convolution is: in, For elements of the two-dimensional convolution result, For input values, Elements of a two-dimensional convolution kernel; Merge and align convolution results: in, These are the calculation results of different convolutions; S3.2.trend Feature Extraction: Removes non-stationary information from the input sequence, especially the mean and standard deviation of instances, and returns the information removed by input normalization to the model to achieve the model's inverse normalization operation. Specifically, the sequence is decomposed into the following parts through the time-series information extraction module: In the formula, For the observation at time t, Represents the trend component at time t; This represents the seasonal component at time t; This represents the residual term at time t; Trend Components The calculation expression is: Seasonal components The calculation expression is: In the formula, , , These represent the y values ​​at times t-1, t-v+1, and tv, respectively. Represents a data set The observation value at the i-th time point, Represents a data set The average of all observations; This represents the average value across all points in time within the dataset. Indicates the length of seasonal cycles; These represent the trend constant and the trend slope, respectively. This represents the time point corresponding to the i-th observation. S4. Public opinion feature extraction: Market public opinion corpus information is converted into word vectors through an embedding layer; After training with the TextCNN model, the encoded results are input into a fully connected layer, and the sentiment features are output through the sigmoid function. S5. Multi-head Attention Calculation: The multi-head attention module performs multi-head attention calculation on the trend features and seasonal features extracted in S3 and the public opinion features extracted in S4, calculating the query, key, and value matrix. Among them, among them, , and This is the weight matrix; further, the attention score is calculated and expressed as: in, It is the dimension of the query and key vector; through Calculate multi-head attention, where It is a weight matrix. It is the output projection matrix; passed through the feedforward neural network layer Processing, among which, It is a weight matrix. , It is the bias vector; S6. Feature Fusion and Prediction: Normalize and perform affine transformation on the three types of output features, and calculate the dynamic weights between each feature. ; in, , These represent seasonal features, public opinion features, and trend features, respectively. It is a learnable projection matrix; These are the mean and standard deviation of the features. These are learnable scaling and offset parameters; through Perform weighted feature fusion, where t represents the time step. This represents the fused feature vector; the fused features are input to the gate control unit for processing, and the gate control is calculated as follows: The state update formula is The gate control output is calculated as follows: in, , , These represent the input gate, forget gate, output gate, and candidate state, respectively. denoted by sigmoid, and tanh by hyperbolic tangent activation function. This represents the state vector from the previous time step. This represents the bias vector; finally, the prediction result is output through a feedforward neural network. ; S7. Prediction Accuracy Assessment: Prediction accuracy is measured using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics, calculated using the following formulas: in, and Let i be the actual value and the predicted value at time i, respectively, and n be the number of data points.

2. The provincial power load forecasting method based on multimodal information fusion analysis according to claim 1, characterized in that: The market data mentioned in S1 includes hourly meteorological data such as temperature, humidity, wind speed, and sunshine, as well as urban economic data, power transmission line data, and holiday schedule information; during data preprocessing, date information is encoded as categorical variables to construct a dataset containing multiple features.

3. The provincial power load forecasting method based on multimodal information fusion analysis according to claim 1, characterized in that: The formula for calculating one-dimensional convolution in S3.1 is as follows: ,in, The i-th element of the convolution result Let be the element at time point t of the input sequence. Let be the ith element of the convolution kernel, and T be the sequence period; the formula for calculating the two-dimensional convolution operation is: ,in, For elements of the two-dimensional convolution result, For input values, The elements of the 2D convolution kernel are used; the result of merging and aligning the convolutions is... , , These are the calculation results of different convolutions.

4. The provincial power load forecasting method based on multimodal information fusion analysis according to claim 1, characterized in that: The public opinion corpus information in S4 is collected through multiple channels and converted into word vectors after preprocessing. The word vectors contain information on positive, negative or neutral market attitudes.

5. The provincial power load forecasting method based on multimodal information fusion analysis according to claim 1, characterized in that: During model training, the number of hidden layer neurons was set to 64, the learning rate was 0.001, and the Adam optimizer was used for parameter updates.

Images