Short-term load prediction optimization method based on feature fusion and ConvLSTM-3D

Abstract

The invention discloses a short-term load prediction optimization method based on feature fusion and ConvLSTM-3D, and relates to the technical field of load prediction, and the method comprises the steps: S1, decomposing historical electrical load data based on time change characteristics, and obtaining load feature components; s2, performing feature screening based on the contribution degree of the load feature component to obtain target features; s3, performing feature fusion on the target features by adopting an attention mechanism to obtain a fused feature sequence; s4, taking the fusion feature sequence as the input of ConvLSTM-3D to carry out model training, and constructing a load prediction optimization model; s5, predicting real-time power consumption data based on the load prediction optimization model, and outputting short-term load prediction data; the problem that the short-term load prediction effect is poor in reliability due to the fact that the influence weight of weather characteristics on the load is not comprehensively considered in the prior art is solved, the short-term load prediction performance and the robustness of the model are remarkably improved, and the accuracy and reliability of the prediction result are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and specifically to a short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D. Background Art

[0002] Short-term load forecasting refers to the forecast of electricity load from one day to one week, which plays an important role in optimizing the allocation of power generation resources and reducing operating costs. Compared with ultra-short-term load forecasting, accurate short-term load forecasting can help power system operators better cope with load changes caused by emergencies (such as extreme weather conditions or sudden social events) and ensure the continuity and safety of power supply. At present, the prediction accuracy of traditional short-term load forecasting models is greatly affected by the structural parameters of the model and the quality of the trained data, and is prone to overfitting or underfitting. There are certain requirements for the integrity and sampling rate of the trained data set. At the same time, the factors affecting the load changes during holidays and the differences in electricity consumption behavior at the user end are not considered enough, resulting in low cross-regional (time and space) load forecasting accuracy.

[0003] Compared with ultra-short-term load forecasting, short-term power load has a relatively long time scale, and its size is affected by time characteristics such as users' power consumption behavior habits, holiday characteristics, and power demand in different time periods. In addition, the degree and size of the impact of different time characteristic components on power consumption are also different. Therefore, how to comprehensively consider the influence weights of specific characteristics such as time characteristics and weather characteristics, and further optimize the processing of short-term load forecasting characteristics is a major challenge in the current load forecasting field.

[0004] Chinese patent, publication number: CN118336721A, publication date: July 12, 2024, discloses a power load forecasting method based on wavelet packet decomposition and ConvLSTM, which uses a wavelet packet decomposition module to decompose the power load data to obtain multiple subsequence data representing different time series rules, and inputs meteorological elements, land use types and subsequence data as multi-source fusion data into the weighted coupling Conv-LSTM network, and uses the weighted coupling module to calculate the channel weighted features to forecast the power load. However, this scheme only decomposes the power load data to obtain subsequence data, and then fuses it with other data for forecasting. It is impossible to know the degree of influence of each data on the forecast result. There may be a fitting phenomenon for the forecast result, and the accuracy of the result cannot be guaranteed, and the forecast effect is poor. Summary of the invention

[0005] The purpose of the present invention is to address the problem that the prior art does not fully consider the impact weight of weather characteristics on the load, resulting in poor reliability of short-term load forecasting effects; a short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D is proposed, by extracting time characteristic components related to the power load, calculating the weight of each component and performing feature screening, and using the attention mechanism to fuse features according to the component weights, and then using the ConvLSTM-3D model to train the fused features to obtain future short-term power load data, thereby improving the short-term load forecasting performance and the robustness of the model, and further improving the accuracy and reliability of the prediction results.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D, comprising the following steps: S1. Decompose the historical power load data based on the time variation characteristics to obtain load characteristic components; S2. Perform feature screening based on the contribution of load feature components to obtain target features; S3, using the attention mechanism to fuse the target features and obtain a fused feature sequence; S4. Use the fused feature sequence as the input of ConvLSTM-3D for model training to build a load forecasting optimization model; S5. Predict the real-time power consumption data based on the load forecasting optimization model and output short-term load forecasting data.

[0007] In this scheme, by decomposing the historical load data, load characteristic components such as trend, seasonality, and holiday effects can be extracted, providing a more accurate and reliable data basis for feature fusion; calculating the contribution of load characteristic components can more clearly understand the influence of each load characteristic component on the prediction model, that is, it can obtain the average contribution of each load characteristic component to the model prediction result. Screening out the target features can simplify the model structure, help improve the robustness and learning ability of the prediction model, and thus improve the prediction efficiency; by fusing the features through the attention mechanism, it can capture the long-term dependencies in the data, and the attention mechanism can automatically learn the weights of different time steps to improve the effect of feature fusion, thereby improving the prediction accuracy; the prediction optimization model is constructed through the ConvLSTM-3D model, which utilizes the characteristics of the ConvLSTM-3D model that can capture the spatiotemporal features in multidimensional data. Compared with the ConvLSTM that can only process three-dimensional data features, the ConvLSTM-3D model is used to train the fused feature sequence, so that the prediction optimization model can learn the complex relationships in the data more comprehensively, thereby improving the accuracy of the prediction results.

[0008] Preferably, S1 comprises the following sub-steps: Based on the Prophet algorithm, different time components of the historical electricity load data are extracted to obtain load characteristic components including trend characteristic components, holiday characteristic components, week characteristic components and daily characteristic components.

[0009] Preferably, S2 comprises the following sub-steps: The load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the contribution of each load characteristic component to the load forecast result; The load feature components are sorted based on the SHAP values, and the load feature components with the highest order are used as target features.

[0010] In this scheme, by calculating the SHAP value of each load characteristic component, the contribution of each characteristic component to the prediction result can be understood, thereby improving the interpretability of the model; feature sorting and screening based on SHAP value helps to optimize the feature selection process, among which, by selecting characteristic components with large contribution to the prediction results, these characteristic components often contain more information related to the prediction target, which can improve the prediction accuracy of the model; removing redundant characteristic components can simplify the model structure, reduce the complexity of the model, help reduce the risk of overfitting of the model, and improve the generalization ability and robustness of the model; in addition, in the feature selection process, removing characteristic components with small contribution to the prediction results can reduce the computational burden of the model and improve the prediction efficiency.

[0011] Preferably, the load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the average contribution of each load characteristic to the load forecast result, including: Calculate the Shapley value of each feature in the trend feature component, the holiday feature component, the week feature component and the day feature component based on the SHAP algorithm; The Shapley values ​​of each feature in the same feature component are added together as the SHAP value of the current feature component.

[0012] As an implementation method, Preferably, S3 includes the following sub-steps: Based on the LSTM algorithm, the target feature is used as input to build a target feature fusion model; Introducing an attention mechanism into the model to calculate the weight coefficient of each target feature; The target features are weightedly fused based on the weight coefficients to obtain the fused feature sequence.

[0013] Preferably, the attention mechanism is introduced into the model to calculate the weight coefficient of each target feature, including: Performing feature encoding on the target features based on model parameters of the target feature fusion model; In the target feature fusion model, an attention mechanism is used to normalize the encoded target features to obtain the weight coefficients corresponding to the target features.

[0014] Preferably, the use of an attention mechanism in the target feature fusion model to normalize the encoded target features and obtain a weight coefficient corresponding to the target features includes: Based on the target feature fusion model that introduces the attention mechanism, the similarity of each target feature is calculated using the temporal hidden state of the target feature and the candidate set; The similarity is input into the activation function for normalization and then used as the weight coefficient of the feature component.

[0015] In this scheme, after the attention mechanism is introduced into the LSTM model, the model can dynamically adjust the weights according to the importance of the input data, so as to better capture the key information in the sequence data; the load data is processed in a weighted summation manner, and attention weights are added to the load features, so that the fusion features cover the load characteristics and change laws, which helps to improve the robustness and generalization ability of the prediction optimization model.

[0016] Preferably, S4 includes the following sub-steps: Based on the input gate of the ConvLSTM-3D network model, the fused feature sequence is trained once to generate first training data and update it into the memory unit; Performing secondary training on the first training data through a forget gate of a ConvLSTM-3D network model to generate second training data; The memory unit state is activated by an activation function and then multiplied with the second training data to determine the output data of the output gate, thereby completing the construction of the load forecasting optimization model.

[0017] Preferably, the input gate based on the ConvLSTM-3D network model trains the fusion feature sequence once, including: The fused feature sequence is input into the input gate, and the input gate performs convolution operation and activation function transformation on the shadow state of the current time step and the previous time step to complete the training.

[0018] Preferably, the generating the first training data and updating it into the memory unit comprises: After the fusion feature sequence and the first training data are processed by the tanh activation function based on the input gate, the processed feature sequence and the first training data are added to the output of the historical memory state of the memory unit to complete the update of the current memory unit state.

[0019] In this solution, the four-dimensional load characteristic data (trend characteristic component, holiday characteristic component, weekly characteristic component and daily characteristic component) extracted by the Prophet algorithm are trained through the ConvLSTM-3D network model to achieve more comprehensive learning of the spatiotemporal characteristics in the data, thereby improving the prediction accuracy. Compared with the traditional ConvLSTM model, which may face the problems of high model complexity and low computational efficiency when processing large-scale data sets, the ConvLSTM-3D model can reduce model complexity and improve computational efficiency while maintaining high prediction accuracy by optimizing the network structure and algorithm, and can better adapt to the load prediction needs in different scenarios, so that the model can maintain high prediction stability and accuracy when facing data in different seasons and different load levels.

[0020] Beneficial effects of the present invention: 1. The four-dimensional load characteristic data (trend characteristic component, holiday characteristic component, week characteristic component and daily characteristic component) extracted by the Prophet algorithm is trained through the ConvLSTM-3D network model to more fully learn the spatiotemporal characteristics in the data, so that the prediction optimization model can more comprehensively learn the complex relationships in the data, thereby improving the accuracy of the prediction results; 2. After introducing the attention mechanism into the LSTM model, the model can dynamically adjust the weights according to the importance of the input data, so as to better capture the key information in the sequence data, which helps to improve the robustness and generalization ability of the prediction optimization model; 3. By calculating the SHAP value of each load feature component, the contribution of each feature component to the prediction result can be understood, thereby improving the interpretability of the model; feature sorting and screening based on SHAP value can help optimize the feature selection process, so as to remove redundant feature components, simplify the model structure, reduce the complexity of the model, help reduce the risk of overfitting of the model, and improve the generalization ability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings. The drawings are only for the purpose of illustrating preferred embodiments and are not to be considered as limiting the present invention. Also, the same reference symbols are used throughout the drawings to represent the same parts.

[0022] Figure 1 This is a flow chart of a short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D in this embodiment.

[0023] Figure 2 This is a statistical diagram of load characteristic component data extracted based on the Prophet algorithm in this embodiment.

[0024] Figure 3 This is a schematic diagram of load characteristic component contribution analysis data based on SHAP value in this embodiment.

[0025] Figure 4 This is a schematic diagram of a feature importance ranking result based on the SHAP mean absolute value in this embodiment.

[0026] Figure 5 A schematic diagram for comparing prediction results of this embodiment.

[0027] Figure 6 It is a schematic diagram for comparing a characteristic component of this embodiment and prediction results in different time periods.

[0028] Figure 7 Schematic diagram of average prediction error distribution of a characteristic component of this embodiment.

[0029] Figure 8 This is a statistical diagram of the prediction results of each prediction model in this embodiment. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0031] Example 1: Figure 1 As shown, a short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D includes steps S1-S5, wherein: S1. Decompose the historical power load data based on the time variation characteristics to obtain load characteristic components.

[0032] Specifically, S1 includes the following sub-steps: Based on the Prophet algorithm, different time components of the historical electricity load data are extracted to obtain load characteristic components including trend characteristic components, holiday characteristic components, week characteristic components and daily characteristic components.

[0033] As an implementation method, the input data of the Prophet algorithm includes the historical electricity load data of the time series. The Prophet algorithm extracts features by integrating holiday information, working day or non-working day information, and the characteristics of the daily electricity load curve, and classifies the extracted feature data according to the above time characteristics. It includes trend feature components, holiday feature components, week feature components, and daily feature components for extraction. Its calculation expression can be expressed as: y(t)=g(t)+s(t)+h(t)+ε(t) (1) Among them, y(t) represents the original electricity load; g(t) represents the trend term, which represents the changing trend of the time series in the non-periodic aspect; s(t) represents the periodic term; h(t) represents the holiday term, which represents the impact of potential non-fixed-periodic holidays in the time series on the predicted value; ε(t) represents the error term, which represents the fluctuation predicted by the model.

[0034] In this embodiment, the Prophet algorithm performs well in processing time series data with complex seasonal factors. It can accurately decompose each characteristic component, automatically identify periodic patterns, consider trend changes, process outliers, and easily interpret and adjust the extracted characteristic components; for example, daily, weekly, and annual seasonal changes can help to fully consider the periodic influencing factors of power load when building a forecasting optimization model. In addition, the Prophet algorithm can automatically identify and process outliers in the data by extracting load characteristic components. For example, the Prophet algorithm can automatically filter out load peaks caused by holidays or special events without preprocessing the power data, simplifying the overall forecasting process, improving the efficiency of feature processing, and avoiding the impact of outliers on the forecasting optimization model, thereby improving forecasting accuracy.

[0035] S2. Perform feature screening based on the contribution of load feature components to obtain target features.

[0036] Specifically, S2 includes the following sub-steps: The load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the contribution of each load characteristic component to the load forecast result; The load feature components are sorted based on the SHAP values, and the load feature components with the highest order are used as target features.

[0037] Specifically, the load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the average contribution of each load characteristic to the load forecast result, including: Calculate the Shapley value of each feature in the trend feature component, the holiday feature component, the week feature component and the day feature component based on the SHAP algorithm; The Shapley values ​​of each feature in the same feature component are added together as the SHAP value of the current feature component.

[0038] As an implementation method, a load model can be characterized as follows: Features p The impact on the output result is related to its corresponding coefficient.

[0039] Features p The incremental contribution of is: The Shapley value is calculated based on the average marginal contribution. And F is the set of all features), for feature i, its shap value is calculated by the following formula: Among them, represents the shap value of feature i, v(S) is a function that represents the predicted output when the model only uses the feature set S, |S| is the number of features in a set S, |F| is the total number of all features, and v(S∪{i})-v(S) represents the marginal contribution when feature i is included compared to when feature i is not included.

[0040] In this embodiment, by calculating the SHAP value of each load characteristic component, the contribution of each characteristic component to the prediction result can be understood, thereby improving the interpretability of the model; feature sorting and screening based on the SHAP value helps to optimize the feature selection process, wherein, by selecting characteristic components that contribute greatly to the prediction results, these characteristic components often contain more information related to the prediction target, which can improve the prediction accuracy of the model; removing redundant characteristic components can simplify the model structure, reduce the complexity of the model, help reduce the risk of overfitting of the model, and improve the generalization ability and robustness of the model; in addition, in the feature selection process, removing characteristic components that contribute little to the prediction results can reduce the computational burden of the model and improve the prediction efficiency.

[0041] S3. Use the attention mechanism to fuse the target features and obtain a fused feature sequence.

[0042] Specifically, S3 includes the following sub-steps: Based on the LSTM algorithm, the target feature is used as input to build a target feature fusion model; Introducing an attention mechanism into the model to calculate the weight coefficient of each target feature; The target features are weightedly fused based on the weight coefficients to obtain the fused feature sequence.

[0043] Specifically, the attention mechanism is introduced into the model to calculate the weight coefficient of each target feature, including: Performing feature encoding on the target features based on model parameters of the target feature fusion model; In the target feature fusion model, an attention mechanism is used to normalize the encoded target features to obtain the weight coefficients corresponding to the target features.

[0044] Specifically, the attention mechanism is used in the target feature fusion model to normalize the encoded target features to obtain the weight coefficient corresponding to the target features, including: Based on the target feature fusion model that introduces the attention mechanism, the similarity of each target feature is calculated using the temporal hidden state of the target feature and the candidate set; The similarity is input into the activation function for normalization and then used as the weight coefficient of the feature component.

[0045] As an implementation method, the fusion feature For example, the similarity scores between the hidden state and different attributes in LSTM are calculated based on the hidden state h at time t-1. t-1 and candidate set h t-1 Calculated and then input into the Softmax function for normalization. The normalized similarity score finally obtains the updated new feature sequence.

[0046] The input components are expressed as follows after feature encoding: Among them, V ε , W εc and W εh These are the parameters obtained based on LSTM structure training.

[0047] The weight coefficients corresponding to different feature components can be expressed as: The feature components reconstructed by the attention mechanism can be expressed as: The fused feature sequence is obtained based on the weighted fusion of the reconstructed feature components, which is expressed as follows: Among them, X kRepresents a fusion feature with k dimensions, It represents the target feature after reconstruction (i.e., with the weight coefficient added), and L represents the time length of the feature component.

[0048] In this embodiment, after the attention mechanism is introduced into the LSTM model, the model can dynamically adjust the weights according to the importance of the input data, so as to better capture the key information in the sequence data; the load data is processed in a weighted summation manner, and attention weights are added to the load features, so that the fused features cover the load characteristics and change laws, which helps to improve the robustness and generalization ability of the prediction optimization model.

[0049] S4. Use the fused feature sequence as the input of ConvLSTM-3D to perform model training and build a load forecasting optimization model.

[0050] Specifically, the S4 includes the following sub-steps: Based on the input gate of the ConvLSTM-3D network model, the fused feature sequence is trained once to generate first training data and update it into the memory unit; Performing secondary training on the first training data through a forget gate of a ConvLSTM-3D network model to generate second training data; The memory unit state is activated by an activation function and then multiplied with the second training data to determine the output data of the output gate, thereby completing the construction of the load forecasting optimization model.

[0051] Specifically, the input gate based on the ConvLSTM-3D network model trains the fusion feature sequence once, including: The fused feature sequence is input into the input gate, and the input gate performs convolution operation and activation function transformation on the shadow state of the current time step and the previous time step to complete the training.

[0052] Specifically, generating the first training data and updating it into the memory unit includes: After the fusion feature sequence and the first training data are processed by the tanh activation function based on the input gate, the processed feature sequence and the first training data are added to the output of the historical memory state of the memory unit to complete the update of the current memory unit state.

[0053] In this embodiment, the four-dimensional load characteristic data (trend characteristic component, holiday characteristic component, week characteristic component and daily characteristic component) extracted by the Prophet algorithm can be trained through the ConvLSTM-3D network model, and the spatiotemporal characteristics in the data can be learned more fully, thereby improving the prediction accuracy. Compared with the traditional ConvLSTM model, which may face the problems of high model complexity and low computational efficiency when processing large-scale data sets, the ConvLSTM-3D model can reduce the model complexity and improve the computational efficiency while maintaining high prediction accuracy by optimizing the network structure and algorithm, and can better adapt to the load prediction requirements in different scenarios, so that the model can maintain high prediction stability and accuracy when facing data in different seasons and different load levels.

[0054] Furthermore, the ConvLSTM-3D model uses ConvLSTM as the basic structural unit and changes the loss function of the traditional ConvLSTM to adapt to the data characteristics based on multi-dimensional feature sequences, as follows: Three-dimensional convolution operations are used in the update of input gate, forget gate, output gate and cell state. The input of each gate of the ConvLSTM unit contains three parts: the memory information of the previous unit, the output of the previous unit and the input at the current moment; among them: Input gate: Determines which parts of the current input should be updated to the memory cell. This process consists of two steps: Determine the information that needs to be updated through the sigmoid layer; Generate alternative information through tanh function; The structure of the input gate can be expressed as: Among them, i t and are the output of the input gate and the backup information of the memory unit, respectively. xt is the input data at the current time t, C t-1 is the state information before the memory unit, H t-1 is the output of the previous hidden layer unit of the ConvLSTM unit, W and b are the weight and bias of the input gate respectively. σ represents the sigmoid activation function, tanh represents the hyperbolic tangent function, and * represents the convolution operation.

[0055] Forget gate: It selectively discards useless information from the memory unit of the previous moment. The forgetting model can be expressed as: f t =σ(W xf * t +W hf *H t-1 +Wcf *C t-1 +b f ) (10) Among them, W xf Indicates the weight of the information from the input layer flowing into the forget gate at this moment, W hf Represents the weight of the final result of the previous hidden layer neural unit at the input of the forget gate, W cf It represents the weight of the memory unit state flowing into the forget gate at the previous moment, and bf represents the bias parameter when the forget gate is calculated.

[0056] Memory unit: The current memory unit state information C is obtained by describing the past long-term state and the current state. t The process of updating the status information can be specifically expressed as follows: Output gate: It consists of two parts. One part is the information input obtained by combining the short-term memory with the current input information (the output of the output gate at the current moment), and the other part is the final output after combining the long-term memory (the output of ConvLSTM at the current moment). The output model can be expressed as: Specifically, in the input gate, the input data passes through the input gate, which receives the input of the current time step (usually a feature map) and the hidden state of the previous time step, and performs convolution operations and activation function transformations on them, usually using the sigmoid function.

[0057] In the forget gate, the forget gate also receives the current input and the hidden state of the previous moment, and generates a "gating" matrix through convolution and sigmoid activation operations, with values ​​between [0,1].

[0058] In the output gate, the current input and the hidden state of the previous time step are received, and then the convolution operation and sigmoid activation are performed. Then, the memory cell state is processed by tanh activation and multiplied with the "gating" matrix of the output gate to determine which information in the hidden state can be output.

[0059] Specifically, the memory state update of the memory unit includes: the old memory state is multiplied by the output of the forget gate to control what is retained and forgotten. The output of the input gate is multiplied with the input information through a candidate memory state (usually processed by the tanh activation function), and then added to the output of the old memory state to update the current memory unit state.

[0060] S5. Predict the real-time power consumption data based on the load prediction optimization model and output short-term load prediction data.

[0061] Beneficial effects of the embodiment: by decomposing historical load data, load characteristic components such as trend, seasonality, and holiday effects can be extracted, providing a more accurate and reliable data basis for feature fusion; calculating the contribution of load characteristic components can more clearly understand the degree of influence of each load characteristic component on the prediction model, that is, the average contribution of each load characteristic component to the model prediction result can be obtained, and screening out target features can simplify the model structure, help improve the robustness and learning ability of the prediction model, and thus improve the prediction efficiency; by fusing features through the attention mechanism, it is possible to capture long-term dependencies in the data, and the attention mechanism can automatically learn the weights of different time steps to improve the effect of feature fusion, thereby improving the prediction accuracy; by constructing a prediction optimization model through the ConvLSTM-3D model, the ConvLSTM-3D model is used to capture the characteristics of the spatiotemporal features in multidimensional data. Compared with the ConvLSTM that can only process three-dimensional data features, the ConvLSTM-3D model is used to train the fused feature sequence, so that the prediction optimization model can learn the complex relationships in the data more comprehensively, thereby improving the accuracy of the prediction results.

[0062] As a specific example analysis, the relevant measured data based on a power grid company in a certain area of ​​Zhejiang Province is used for verification, wherein the load data part records the power load of some areas in the region from 2021 to 2023, and the sampling rate of the relevant data is one sampling point every 15 minutes. The weather data records the surface temperature, wind speed, wind direction, humidity and other weather data in different areas of the region, and has the same sampling rate; the important information about the load and meteorological data of the region is shown in Table 1, Table 1 Important information of the measured data set This embodiment selects the total social load data of a city in the region as verification data, of which 80% of the data is used for the training set and the remaining 20% ​​is used for the test set; In order to verify the robustness of the load forecasting optimization model, the load data of the other three regions are used as supplementary verification, and the meteorological data of the corresponding regions are used. The time feature extraction results of this experimental example are as follows: Figure 2 As shown in the figure, the characteristic component contribution analysis results of SHAP value in this experimental example are as follows Figure 3 As shown in the figure, the ranking results of feature importance in this experimental example are as follows Figure 4 shown.

[0063] Furthermore, the feature quantity and prediction model are compared and verified through ablation experiments. In terms of feature quantity, the comparative analysis uses a single time feature component, a single weather feature component, and a fusion feature sequence for comparison; in terms of prediction model, LSTM, CNN-LSTM, ConvLSTM, and ConvLSTM-3D used in this embodiment are compared and analyzed; among them, according to whether different time periods contain holiday features, the power load data during the National Day from May 1st to May 3rd is selected, and the power load data of the area during non-holiday periods is compared. The prediction results are as follows: Figure 5 As shown. Through comparative analysis, it can be concluded that due to the impact of holidays, the power load in this area shows a certain growth trend. And the load forecasting optimization model adopted in this embodiment shows the best forecasting performance both on holidays and non-holidays.

[0064] Furthermore, by using the time features based on Prophet feature components, the weather features based on weather feature components, and the fusion features of time-weather components for comparative analysis, this experimental example compares the load forecast results during the non-holiday period (April 1st to April 3rd) and the holiday period (May 1st to May 3rd). The forecast result curve is shown in Figure 6 As shown in the figure, the average prediction error of the two prediction periods is Figure 7 As shown in Table 2, the error distribution of prediction using different feature components can be further analyzed. The experimental results of the ablation based on feature quantity analysis are shown in Table 2. Table 2 Ablation experimental results based on feature quantity analysis Furthermore, a five-fold cross validation was used to verify the prediction results, where the verified data were the load data of regions 2 and 3 and the corresponding meteorological data. The prediction results are shown in Table 3. Table 3 Test results based on five-fold cross validation Considering the differences between different devices in different regions, the downsampling method is used to increase the prediction performance of the model at different sampling rates. The results are shown in Table 4. Table 4 Test results based on five-fold cross validation Based on the above experimental data, the prediction performance of the model is statistically analyzed, and the statistical results are as follows: Figure 8In summary, the indicators of the short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D neural network are at a high level, and the impact of holiday load changes and user-side electricity consumption behavior differences is deeply considered, which improves the short-term load forecasting performance and the robustness of the model. It overcomes the problem of poor reliability of short-term load forecasting results due to the failure to fully consider the impact weight of weather characteristics on load in the existing technology, and significantly improves the accuracy and reliability of the forecast results.

[0065] The above specific embodiments are preferred embodiments of the present invention, and are not intended to limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to the present specific embodiments. All equivalent changes made in accordance with the shape, structure, and method of the present invention are within the protection scope of the present invention.

Claims

1. A short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D, characterized by: The steps include: S1. Decompose the historical power load data based on the time variation characteristics to obtain load characteristic components; S2. Perform feature screening based on the contribution of load feature components to obtain target features; S3, using the attention mechanism to fuse the target features and obtain a fused feature sequence; S4, using the fused feature sequence as the input of ConvLSTM-3D to perform model training and build a load forecasting optimization model; S5. Predict the real-time power consumption data based on the load prediction optimization model and output short-term load prediction data.

2. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 1 is characterized in that: The S1 comprises the following sub-steps: Based on the Prophet algorithm, different time components of the historical electricity load data are extracted to obtain load characteristic components including trend characteristic components, holiday characteristic components, week characteristic components and daily characteristic components.

3. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 2 is characterized in that: The S2 comprises the following sub-steps: The load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the contribution of each load characteristic component to the load forecast result; The load feature components are sorted based on the SHAP values, and the load feature components with the highest order are used as target features.

4. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 3 is characterized in that: The load characteristic components are calculated based on the SHAP algorithm to obtain a SHAP value representing the average contribution of each load characteristic component to the load forecast result, including: Calculate the Shapley value of each feature in the trend feature component, the holiday feature component, the week feature component and the day feature component based on the SHAP algorithm; The Shapley values ​​of each feature in the same feature component are added together as the SHAP value of the current feature component.

5. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 1 is characterized in that: The S3 comprises the following sub-steps: Based on the LSTM algorithm, the target features are used as input to build a target feature fusion model; Introducing an attention mechanism into the model to calculate the weight coefficient of each target feature; The target features are weightedly fused based on the weight coefficients to obtain the fused feature sequence.

6. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 5 is characterized in that: The introducing of the attention mechanism into the model to calculate the weight coefficient of each target feature includes: performing feature encoding on the target feature based on the model parameters of the target feature fusion model; In the target feature fusion model, an attention mechanism is used to normalize the encoded target features to obtain the weight coefficients corresponding to the target features.

7. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 6 is characterized in that: The method of using the attention mechanism in the target feature fusion model to normalize the encoded target features and obtain the weight coefficients corresponding to the target features includes: Based on the target feature fusion model that introduces the attention mechanism, the similarity of each target feature is calculated using the temporal hidden state of the target feature and the candidate set; The similarity is input into the activation function for normalization and then used as the weight coefficient of the feature component.

8. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 1 is characterized in that: The S4 comprises the following sub-steps: Based on the input gate of the ConvLSTM-3D network model, the fused feature sequence is trained once to generate first training data and update it into the memory unit; Performing secondary training on the first training data through a forget gate of a ConvLSTM-3D network model to generate second training data; The memory unit state is activated by an activation function and then multiplied with the second training data to determine the output data of the output gate, thereby completing the construction of the load forecasting optimization model.

9. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 8 is characterized in that: The input gate based on the ConvLSTM-3D network model performs one training on the fused feature sequence, including: inputting the fused feature sequence into the input gate, and the input gate performs convolution operation and activation function transformation on the hidden state of the current time step and the previous time step to complete the one training.

10. The short-term load forecasting optimization method based on feature fusion and ConvLSTM-3D according to claim 8 is characterized in that: The generating of the first training data and updating the first training data into the memory unit comprises: After the fusion feature sequence and the first training data are processed by the tanh activation function based on the input gate, the result is added to the output of the historical memory state of the memory unit to complete the update of the current memory unit state.

Images