Short-term power load prediction method and device based on adaptive deep fusion

Abstract

The present application relates to a kind of short-term power load prediction method and equipment based on adaptive deep fusion, it is related to power transmission and transformation equipment detection relevant technical field, including: obtaining multiple feature power load data, standardization is handled to obtain sequence window data;Perform statistical analysis based on the instantaneous entropy of multiple feature dimensions, modeling is multimodal instantaneous entropy feature vector;Power prediction module is built;Sequence window data is input into the first branch of double-branch prediction network, and multimodal instantaneous entropy feature vector is input into second branch, and time series features and entropy features are output in parallel, feature fusion and short-term power load prediction are executed, determine load prediction value.The technical problems of insufficient feature utilization, modal fusion mechanism rigidity and insufficient description of load sequence instantaneous dynamic characteristics in the prior art are solved, the technical effects of deeply mining the internal correlation of multiple features, dynamically fusing multimodal information and significantly improving the prediction accuracy and robustness of complex fluctuation load are achieved.

Description

Technical Field

[0001] This application relates to the technical field of power transmission and transformation equipment testing, specifically to a method and equipment for short-term power load forecasting based on adaptive deep fusion. Background Technology

[0002] Short-term load forecasting (typically referring to forecasts for the next few hours to days) is a core basis for power system dispatching, trading plan formulation, unit allocation, and operation control. Its accuracy directly affects the power grid's supply and demand balance, operating costs, and power supply security. Traditional short-term load forecasting methods mainly rely on historical load time series, employing statistical models (such as autoregressive integral moving average models) or early machine learning methods (such as support vector machines). Although structurally simple, these methods have limited ability to characterize the complex nonlinear and non-stationary characteristics of load series, as well as the coupling effects of various external influencing factors (such as weather, date type, and holidays).

[0003] Preferably, in recent years, deep learning, represented by Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs) and their variants, has made significant progress in this field. However, existing models based on single deep networks often have many limitations. Electricity load is typically driven by multiple factors. Existing methods introduce multi-dimensional features (such as temperature, humidity, and date), and usually perform simple concatenation or weighting, failing to deeply explore the dynamic and non-linear correlation patterns between different feature dimensions. In particular, they fail to effectively quantify and utilize the complexity and uncertainty information exhibited by the load sequence and its related features in an instant or short period of time. When multi-branch networks process different types of features separately, they generally use pre-set fixed weights or simple concatenation / addition methods for feature fusion. This makes it difficult to adapt to the time-varying nature of the contribution of different features during dynamic load changes, resulting in low information fusion efficiency and decreased prediction performance when load fluctuations are severe or special events occur. Load changes not only exhibit macro-trends and periodicity, but their micro-instantaneous fluctuations also contain important information. Conventional time series models focus on the long-term dependencies of sequences, but have weak ability to conduct targeted analysis and feature extraction of abrupt changes, noise, and irregular patterns in short-term sequences, affecting the model's response speed and prediction accuracy to rapidly changing loads.

[0004] Therefore, current technologies suffer from problems such as insufficient utilization of features, rigid modal fusion mechanisms, and inadequate characterization of the instantaneous dynamic characteristics of load sequences. Summary of the Invention

[0005] This application provides a method and device for short-term power load forecasting based on adaptive deep fusion, which solves the technical problems of insufficient feature utilization, rigid modal fusion mechanism and insufficient characterization of instantaneous dynamic characteristics of load sequence in the prior art. It achieves the technical effect of deeply mining the intrinsic correlation of multiple features, dynamically fusing multimodal information and significantly improving the accuracy and robustness of forecasting complex fluctuating loads.

[0006] This application provides a short-term power load forecasting method based on adaptive deep fusion. The method includes: acquiring multi-feature power load data, performing standardization processing to obtain sequence window data, wherein the standardization method includes preprocessing and instantaneous segmentation; performing statistical analysis on each sequence window data based on instantaneous entropy of multiple feature dimensions, modeling it as a multimodal instantaneous entropy feature vector; constructing a power forecasting module, wherein the power forecasting module includes a dual-branch prediction network, a fusion network, and a prediction network; inputting the sequence window data into the first branch of the dual-branch prediction network, inputting the multimodal instantaneous entropy feature vector into the second branch, outputting time-series features and entropy features in parallel, performing feature fusion based on the fusion network and short-term power load forecasting based on the prediction network, and determining the load forecast value.

[0007] In a possible implementation, the short-term power load forecasting method based on adaptive deep fusion includes: the instantaneous entropy includes at least approximate entropy, sample entropy and permutation entropy; the statistical features include at least mean, standard deviation and median.

[0008] In a possible implementation, the short-term power load forecasting method based on adaptive deep fusion further includes the following parameters: approximate entropy, sample entropy, and permutation entropy are respectively: approximate entropy m=2, r=0.2×sequence standard deviation; sample entropy m=2, r=0.2×sequence standard deviation; permutation entropy m=3, τ=1, where m is the embedding dimension, r is the similarity tolerance coefficient, and τ is the time delay.

[0009] In a possible implementation, the short-term power load forecasting method based on adaptive deep fusion further performs the following processing: the first branch is trained and constructed using a long short-term memory network for feature extraction, and the second branch is trained and constructed using a multilayer perceptron for feature transformation; the fusion network is trained and constructed using a hierarchical fusion mechanism; the prediction network uses the Huber loss function or the mean squared error loss function, and the optimizer uses Adam or AdamW.

[0010] In a possible implementation, the short-term power load forecasting method based on adaptive deep fusion further performs the following processing: the hierarchical fusion mechanism includes a first fusion mechanism and a second fusion mechanism; if the forecasting task is a load forecasting type, the first fusion mechanism is activated, wherein the first fusion mechanism is feature splicing fusion; if the forecasting task is a cross-modal interaction type, the second fusion mechanism is activated, wherein the second fusion mechanism is multi-head sub-attention fusion.

[0011] In a possible implementation, standardization processing is performed to obtain sequence window data, including: identifying the multi-feature power load data, performing missing value identification and forward imputation processing to determine preprocessed data; and dividing the preprocessed data into time-series-based sliding window segments by presetting a first window length to determine the sequence window data.

[0012] In a possible implementation, the short-term power load forecasting method based on adaptive deep fusion further performs the following processing: After determining the load forecast value, the evaluation of the forecast result adopts a dynamic accuracy standard based on the data range, and the calculation formula is as follows: ; in: ; in, The prediction error is The proportion of sample data below The weights are set to be within 5% of the prediction error. The weight values ​​are those with a prediction error within 10%. The weights are defined as follows: N represents the total number of samples, and I(·) is the indicator function. For the first A sequence of true values ​​for a sample data set. For the corresponding time window The predicted value sequence, The error between the maximum and minimum values ​​represents the normalization error. This represents the percentage of error.

[0013] This application also provides an electronic device, including: a memory for storing executable instructions; and a processor for implementing a short-term power load forecasting method based on adaptive deep fusion when executing the executable instructions stored in the memory.

[0014] This application proposes a method and device for short-term power load forecasting based on adaptive deep fusion. The method acquires multi-feature power load data, performs standardization to obtain sequence window data, performs statistical analysis based on instantaneous entropy of multiple feature dimensions, and models it as a multimodal instantaneous entropy feature vector. A power forecasting module is constructed. The sequence window data is input into the first branch of a dual-branch forecasting network, and the multimodal instantaneous entropy feature vector is input into the second branch. The time-series features and entropy features are output in parallel. Feature fusion and short-term power load forecasting are then performed to determine the predicted load value. This method solves the technical problems of insufficient feature utilization, rigid modal fusion mechanisms, and insufficient characterization of the instantaneous dynamic characteristics of load sequences in existing technologies. It achieves the technical effects of deeply mining the intrinsic correlation of multiple features, dynamically fusing multimodal information, and significantly improving the accuracy and robustness of forecasting complex fluctuating loads. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0016] Figure 1 A schematic diagram of the short-term power load forecasting method based on adaptive deep fusion provided in the embodiments of this application.

[0017] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0018] Explanation of reference numerals in the attached drawings: Input device 401, processor 402, memory 403, output device 404. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0020] This application provides a short-term power load forecasting method based on adaptive deep fusion, such as... Figure 1 As shown, the method includes: Step S100: Obtain multi-feature power load data and perform standardization processing to obtain sequence window data. The standardization method includes preprocessing and instantaneous segmentation.

[0021] Preferably, a SCADA system is used for monitoring, control, and data acquisition of the wide-area power grid, obtaining high-precision, high-frequency (second-level / minute-level) data such as regional total load, substation load, and power flow of critical lines. Simultaneously, smart meters are used to remotely and automatically collect and transmit user-side (residential, industrial, and commercial) electricity consumption data, obtaining detailed load data on the user side. Furthermore, sequence data such as temperature, humidity, wind speed, solar radiation intensity, and perceived temperature are obtained from meteorological bureaus, commercial meteorological APIs, or proprietary meteorological stations, along with electricity price information, announcements of major events, and calendar data such as date type (weekday, weekend, statutory holiday) and season, forming multi-feature power load data. This data is then standardized, including preprocessing and instantaneous segmentation. Preprocessing addresses quality issues in the original data, ensuring data integrity and consistency, including the identification and supplementation of missing values. Instantaneous segmentation constructs a sample set with a fixed time span from the continuous preprocessed data time stream, i.e., using a time-series-based sliding window segmentation through a preset window, outputting multiple fixed-shape sequence window data.

[0022] Furthermore, step S100 also includes identifying the multi-feature power load data, performing missing value identification and forward filling processing to determine preprocessed data; and performing time-series-based sliding window segmentation on the preprocessed data by setting a first window length to determine the sequence window data.

[0023] Preferably, the system identifies multi-feature power load data, systematically scans the data, locates numerical gaps caused by sensor failures, transmission interruptions, etc., and fills the identified missing values ​​using forward padding, that is, using the observation value of the previous valid moment before the missing position to fill the current missing value, outputting preprocessed data that is continuous and without missing values ​​in the time dimension. Then, a fixed time step is preset as the first window length, such as 24 hours, which means using the data of the past 24 hours to predict the next hour. Based on the first window length, starting from the starting time point of the preprocessed data, the system slides sequentially in time order. The first sample takes all feature data from time point 1 to the first window length as input features. The second sample takes data from time point 2 to the first window length + 1 as input features after sliding the window forward one time step, and so on, until the end of the data, outputting multiple fixed-shape sequence window data.

[0024] Step S200: Perform statistical analysis on the instantaneous entropy based on multiple feature dimensions on the data of each sequence window, and model it as a multimodal instantaneous entropy feature vector.

[0025] Preferably, statistical analysis of instantaneous entropy is performed on each feature dimension in each sequence window data. This involves splitting the sequence window data into multiple one-dimensional time series, representing load series, temperature series, humidity series, etc., to quantify the dynamic complexity of each influencing factor within the window period. Instantaneous entropy refers to the entropy value calculated immediately using data within a specific time window, used to characterize the regularity or randomness of the sequence within that short period. For each one-dimensional sequence, multiple entropy algorithms are used to calculate and determine multiple entropy values ​​for each feature sequence, including approximate entropy value, sample entropy value, and permutation entropy value. Furthermore, the statistical characteristics of the sequence within the window, such as mean, standard deviation, and median, are calculated. Then, the entropy value and statistical characteristics of each feature sequence are constructed into a 6-dimensional feature sub-vector. Finally, each sequence window data is modeled as a highly condensed multimodal instantaneous entropy feature vector that characterizes the complexity of the short-term system state, thereby improving the model's ability to characterize complex fluctuations.

[0026] Furthermore, step S200 also includes that the instantaneous entropy includes at least approximate entropy, sample entropy, and permutation entropy; and the statistical features include at least mean, standard deviation, and median.

[0027] Preferably, the instantaneous entropy includes at least approximate entropy, sample entropy, and permutation entropy. Approximate entropy measures the probability of a sequence generating a new pattern; a larger value indicates greater complexity. Sample entropy is an improvement on approximate entropy, with lower dependence on data length. Permutation entropy measures complexity by comparing the order of sequence values, offering high computational efficiency and sensitivity to dynamic mutations. Statistical features include at least the mean, standard deviation, and median, which quantify the average level of the sequence within the window, the degree of fluctuation of the sequence within the window, and the central tendency of the sequence within the window, respectively, and are insensitive to outliers.

[0028] Furthermore, step S200 also includes the following parameters for approximate entropy, sample entropy and permutation entropy: approximate entropy m=2, r=0.2×sequence standard deviation; sample entropy m=2, r=0.2×sequence standard deviation; permutation entropy m=3, τ=1, where m is the embedding dimension, r is the similarity tolerance coefficient and τ is the time delay.

[0029] Preferably, the parameters for calculating approximate entropy, sample entropy, and permutation entropy include embedding dimension, similarity tolerance coefficient, and time lag. The embedding dimension represents the number of consecutive data points used to construct a "state" or "pattern" when calculating entropy, determining the level of detail in examining the sequence pattern. m=2 indicates that the pattern formed by examining the size arrangement of two consecutive points is examined. The similarity tolerance coefficient is the distance threshold for judging whether two patterns are similar. If the sequence fluctuates greatly, the similarity standard is relaxed; if the sequence is stable, the standard is tightened to accommodate data with different levels of fluctuation. The time lag is the time interval for sampling data points when constructing the permutation pattern. τ=1 indicates that consecutive adjacent data points are used to examine the permutation order.

[0030] Step S300: Construct a power prediction module, wherein the power prediction module includes a dual-branch prediction network, a fusion network, and a prediction network.

[0031] Preferably, a power forecasting module based on a neural network architecture is constructed for short-term power load forecasting tasks. This module includes a dual-branch forecasting network, a fusion network, and a forecasting network. Specifically, the first branch of the dual-branch forecasting network uses a Long Short-Term Memory (LSTM) network to process sequence window data. It inputs multi-dimensional time series windows containing hourly load, temperature, and humidity data from the past 24 hours. The LSTM, through its gating mechanism (forget gate, input gate, output gate), learns the long-term and short-term dependencies, periodic trends, and dynamic evolution patterns in the input sequence, outputting one or more hidden state vectors to determine the "time series features" of the original sequence after deep abstraction. The second branch of the dual-branch forecasting network uses a Multilayer Perceptron (MLP) to process multimodal instantaneous entropy feature vectors. These vectors are input and, through multiple nonlinear fully connected layers, learn the complex nonlinear combinations and interactions between these entropy values ​​and statistics, mapping the original entropy feature vectors to a higher-level, more abstract representation called "entropy features." The outputs of the two branches are two sets of feature vectors generated simultaneously but with different feature spaces, together forming a dual description of the system state at the current forecast time.

[0032] Preferably, the fusion network is used to achieve information complementarity and adaptive integration. A hierarchical fusion mechanism is adopted to effectively combine the "temporal features" of the first branch and the "entropy features" of the second branch. Specifically, the first fusion mechanism is feature concatenation fusion, which is activated when the prediction task is identified as a regular load prediction class, and directly connects the output feature vectors of the two branches. The second fusion mechanism is multi-head sub-attention fusion, which is activated when the prediction task requires stronger cross-modal interaction, such as when there are complex event interferences or extreme weather. Using the multi-head attention mechanism, the "temporal features" and "entropy features" can perform interactive queries, key calculations, and value calculations, so that the model can dynamically focus on the part of the two features that is most relevant to the current context, achieving more refined and adaptive feature fusion, and finally outputting a fused feature representation.

[0033] Furthermore, step S300 also includes: the first branch is trained and constructed using a long short-term memory network for feature extraction; the second branch is trained and constructed using a multilayer perceptron for feature transformation; the fusion network is trained and constructed using a hierarchical fusion mechanism; the prediction network uses the Huber loss function or the mean squared error loss function; and the optimizer uses Adam or AdamW.

[0034] Preferably, the prediction network is trained using a hierarchical fusion mechanism, that is, a multilayer perceptron with fewer layers performs the final load value regression prediction. The input is the feature vector output by the fusion network, and Huber loss or mean squared error is used as the loss function. Huber loss is a combination of mean squared error and mean absolute error, which can improve the robustness of the model. Mean squared error is the standard loss for regression tasks, which emphasizes the squared penalty of error. The optimizer uses Adam or AdamW, which can train deep networks efficiently and stably. AdamW can better prevent overfitting. The final output is the load prediction value, such as the load for the next hour or the load sequence for the next many hours.

[0035] Furthermore, step S300 also includes the following: the hierarchical fusion mechanism includes a first fusion mechanism and a second fusion mechanism; if the prediction task is a load prediction type, the first fusion mechanism is activated, wherein the first fusion mechanism is feature splicing fusion; if the prediction task is a cross-modal interaction type, the second fusion mechanism is activated, wherein the second fusion mechanism is multi-head sub-attention fusion.

[0036] Preferably, the hierarchical fusion mechanism is an intelligent fusion mechanism that adaptively selects fusion strategies based on task type. It includes a first fusion mechanism and a second fusion mechanism. By using a predefined task classifier or heuristic rules to determine whether the current prediction task belongs to the "load prediction class" or the "cross-modal interaction class", the corresponding fusion mechanism is automatically invoked. Specifically, if the prediction task is determined to be "load prediction class", it is usually a regular prediction scenario with strong regularity, relatively stable patterns, and few external sudden interferences. For example, short-term load prediction is performed on weekdays with stable weather and no special holidays. Then, the first fusion mechanism is activated, and the temporal feature vector output by the first branch LSTM and the entropy feature vector output by the second branch MLP are directly connected in the feature dimension. The decision-making power of feature selection and weight allocation is completely handed over to the prediction network of the fully connected layer. Through training, different importance weights are automatically learned to be assigned to different dimensions.

[0037] Preferably, if the prediction task is determined to be "cross-modal interaction type", which usually refers to scenarios with complex situations, strong nonlinear interference, or the need for a deep understanding of the relationship between features, such as during extreme weather events (such as heat waves and cold waves), the relationship between temperature and load changes drastically, and major holidays or emergencies cause the load pattern to deviate completely from the daily pattern, then the second fusion mechanism is activated to perform feature fusion. Specifically, the attention mechanism is used to dynamically calculate the fusion weights and capture complex dependencies across features. Attention may be applied to feature subsets representing different "modalities" composed of two branch features, and multiple independent attention calculations are run in parallel. Each head learns the relationship between features in a different representation subspace, and finally the results are merged to enhance the model's ability to capture multiple different types of dependencies.

[0038] Step S400: Input the sequence window data into the first branch of the dual-branch prediction network, input the multimodal instantaneous entropy feature vector into the second branch, output the time series features and entropy features in parallel, perform feature fusion based on the fusion network and short-term power load prediction based on the prediction network, and determine the load prediction value.

[0039] Preferably, the sequence window data is input into the first branch of the dual-branch prediction network. The LSTM network reads the multi-dimensional feature vectors of each time point sequentially, and learns the temporal dependencies, trends, and periodic patterns in the sequence through its internal gating mechanism. The hidden state of the last time step of the LSTM is taken, or the hidden states of all time steps are aggregated. For example, average pooling is used to output a high-dimensional vector-temporal feature, which is a deep abstraction and semantic encoding of the entire input sequence. The multimodal instantaneous entropy feature vector is input into the second branch. The multilayer perceptron, composed of multiple fully connected layers and nonlinear activation functions, maps the input multi-dimensional vector to a new feature space through nonlinear transformation, and also outputs a high-dimensional vector-entropy feature. According to the hierarchical fusion mechanism, the temporal feature vector output by the first branch and the entropy feature vector output by the second branch are fused to obtain a fused vector, which is then input into the prediction network for short-term power load prediction. That is, the high-dimensional fused feature is mapped to a specific value as the final load prediction value. Among them, single-step prediction outputs the predicted load of the next time step, and multi-step prediction outputs the predicted load sequence of multiple time steps.

[0040] Furthermore, step S400 also includes, after determining the load forecast value, evaluating the forecast result using a dynamic accuracy standard based on the data range, calculated using the following formula: ; in: ; in, The prediction error is The proportion of sample data below The weights are set to be within 5% of the prediction error. The weight values ​​are those with a prediction error within 10%. The weights are defined as follows: N represents the total number of samples, and I(·) is the indicator function. For the first A sequence of true values ​​for a sample data set. For the corresponding time window The predicted value sequence, The error between the maximum and minimum values ​​represents the normalization error. This represents the percentage of error.

[0041] The short-term power load forecasting system based on adaptive deep fusion provided in this embodiment of the invention can execute the short-term power load forecasting method based on adaptive deep fusion provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0042] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, showing a block diagram of an exemplary electronic device suitable for implementing the embodiments of the present invention. Figure 2 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention. This electronic device is in the form of a general-purpose computing device, and its components may include, but are not limited to, an input device 401, a processor 402, a memory 403, and an output device 404. The processor 402 may be one or more; the memory 403 may include a computer-readable medium and at least one program product having a set (at least one) of program modules configured to perform the functions of the embodiments of this application.

[0043] The memory 403 shown in this embodiment of the invention can be any combination of one or more computer-readable media. The computer-readable storage media can be, but is not limited to, infrared, semiconductor systems, devices or components, or any combination thereof, used to store software programs, computer-executable programs and modules, such as the program instructions / modules corresponding to the short-term power load forecasting method based on adaptive deep fusion in this embodiment of the invention. The processor 402 executes various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory 403, thereby realizing the above-mentioned short-term power load forecasting method based on adaptive deep fusion.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A short-term power load forecasting method based on adaptive deep fusion, characterized in that, The method includes: Multi-feature power load data is acquired, and standardized processing is performed to obtain sequence window data. The standardization method includes preprocessing and instantaneous segmentation. Statistical analysis based on multi-dimensional feature dimensions of instantaneous entropy was performed on the data of each sequence window, and the model was formed as a multimodal instantaneous entropy feature vector; A power prediction module is constructed, wherein the power prediction module includes a dual-branch prediction network, a fusion network, and a prediction network; The sequence window data is input into the first branch of the dual-branch prediction network, and the multimodal instantaneous entropy feature vector is input into the second branch. The time series features and entropy features are output in parallel. Feature fusion based on the fusion network and short-term power load prediction based on the prediction network are performed to determine the load prediction value.

2. The short-term power load forecasting method based on adaptive deep fusion as described in claim 1, characterized in that, The instantaneous entropy includes at least approximate entropy, sample entropy, and permutation entropy; Statistical characteristics include at least the mean, standard deviation, and median.

3. The short-term power load forecasting method based on adaptive deep fusion as described in claim 2, characterized in that, The parameters for approximate entropy, sample entropy, and permutation entropy are as follows: approximate entropy m=2, r=0.2×sequence standard deviation; sample entropy m=2, r=0.2×sequence standard deviation; permutation entropy m=3, τ=1, where m is the embedding dimension, r is the similarity tolerance coefficient, and τ is the time delay.

4. The short-term power load forecasting method based on adaptive deep fusion as described in claim 1, characterized in that, The first branch is built using a long short-term memory network for feature extraction training, and the second branch is built using a multilayer perceptron for feature transformation training. The fusion network is trained and built using a hierarchical fusion mechanism; The prediction network uses either the Huber loss function or the mean squared error loss function, and the optimizer uses either Adam or AdamW.

5. The short-term power load forecasting method based on adaptive deep fusion as described in claim 4, characterized in that, The hierarchical fusion mechanism includes a first fusion mechanism and a second fusion mechanism; If the prediction task is a load prediction type, the first fusion mechanism is activated, wherein the first fusion mechanism is feature splicing fusion; If the prediction task is a cross-modal interaction type, the second fusion mechanism is activated, wherein the second fusion mechanism is multi-head sub-attention fusion.

6. The short-term power load forecasting method based on adaptive deep fusion as described in claim 1, characterized in that, The normalization process yields sequence window data, including: The multi-feature power load data is identified, missing value identification and forward imputation are performed, and preprocessed data is determined. By presetting a first window length, the preprocessed data is divided into time-series sliding window segments to determine the sequence window data.

7. The short-term power load forecasting method based on adaptive deep fusion as described in claim 1, characterized in that, After determining the load forecast values, the evaluation of the forecast results adopts a dynamic accuracy standard based on the data range, and the calculation formula is as follows: ; in: ; in, The prediction error is The proportion of sample data below The weights are set to be within 5% of the prediction error. The weight values ​​are those with a prediction error within 10%. The weights are defined as follows: N represents the total number of samples, and I(·) is the indicator function. For the first A sequence of true values ​​for a sample data set. For the corresponding time window The predicted value sequence, The error between the maximum and minimum values ​​represents the normalization error. This represents the percentage of error.

8. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the short-term power load forecasting method based on adaptive deep fusion as described in any one of claims 1-7.

Images