Power load prediction method and system based on liquid neural network and multilevel interpretable module
By combining liquid neural networks with multi-level interpretable modules, the problems of interpretability, non-stationarity, and robustness in power load forecasting are solved, achieving high-precision and transparent power load forecasting and supporting real-time decision-making in power grid dispatching.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HEBEI ELECTRIC POWER CO LTD
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-15
AI Technical Summary
Existing power load forecasting methods suffer from insufficient model interpretability, limited ability to model non-stationarity, and insufficient robustness under abrupt change scenarios, which affect the accuracy and reliability of power grid dispatching.
By employing a liquid neural network combined with a time attention mechanism and a multi-level interpretable module, the non-stationary characteristics of power load are modeled through adaptive time constants and sparse connections, and multi-level interpretability analysis is performed to improve prediction accuracy and transparency.
It improves the accuracy of power load forecasting and the transparency of the model, enhances robustness in sudden change scenarios, and supports real-time decision-making for power grid dispatch.
Smart Images
Figure CN122051940A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power load forecasting, specifically relating to a power load forecasting method and system based on liquid neural networks and multi-level interpretable modules. Background Technology
[0002] Electricity load forecasting is a crucial aspect of power system operation and planning, and its accuracy directly impacts power dispatch, generation planning, and the safety and economy of grid operation. Traditional load forecasting methods primarily include time series analysis methods (such as ARIMA and seasonal index models), statistical regression models, and machine learning-based forecasting methods. While these methods are effective in modeling linear patterns and stable trends, their predictive performance is poor for the non-stationarity, nonlinearity, and sudden fluctuations prevalent in electricity load data.
[0003] In recent years, deep learning methods have been widely used in load forecasting, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and attention mechanisms. These methods have advantages in modeling complex time-series patterns, but they still have the following shortcomings. First, the models lack interpretability. As black-box models, deep neural networks, while improving prediction accuracy, lack transparent interpretation of the prediction results, which is detrimental to the trust and decision support of power grid dispatchers. Second, their ability to model non-stationarity is limited. Traditional RNNs or LSTMs rely on fixed time constants, making it difficult to effectively model dynamically changing load processes. Furthermore, the models lack robustness to abrupt changes. Power load fluctuates drastically during holidays, extreme weather, and sudden events, and the prediction performance of existing models deteriorates significantly under these abrupt changes.
[0004] Current electricity load forecasting methods face the following challenges: (1) Insufficient interpretability of the model: Most of the existing deep neural networks are “black box” models. Although they can improve the accuracy of load forecasting, their internal reasoning mechanism is difficult to present intuitively and lacks a clear explanation of the forecast results, which affects the power grid dispatchers’ trust in the model and its application value in actual decision-making.
[0005] (2) Limited ability to model nonstationary processes: Traditional recurrent neural networks (RNN) and long short-term memory networks (LSTM) generally use fixed time constants, which makes it difficult to flexibly characterize the dynamic changes of load data at different time scales, resulting in limited modeling effect on nonstationary processes.
[0006] (3) Insufficient robustness under sudden change scenarios: During holidays, extreme weather conditions or sudden events, power load often fluctuates drastically. Existing methods often have significant prediction bias under such sudden change scenarios, showing poor stability and adaptability. Summary of the Invention
[0007] In view of this, the purpose of this invention is to overcome the shortcomings of existing technologies, such as lack of interpretability, limited ability to extract non-stationary features, and poor application robustness. This invention proposes a power load forecasting method and system based on liquid neural networks and multi-level interpretable modules. By introducing liquid neural networks to model the dynamic non-stationary characteristics of load data, and combining time attention mechanisms and multi-level interpretable modules, the method improves forecast accuracy and makes forecast results transparent. At the same time, it adapts to various application scenarios and provides a reliable basis for power grid scheduling and operation.
[0008] Based on the above-mentioned technical problems, the present invention adopts the following technical solution: Collect multi-source heterogeneous data from the power system, and preprocess the multi-source heterogeneous data to construct initial input features; The importance of different initial input features is calculated based on a gating mechanism and dynamically weighted to obtain a weighted residual feature representation. The weighted residual feature representation is extracted using a liquid neural network. The non-stationary dynamic characteristics of the power load are modeled by liquid neurons with adaptive time constants and sparse connections, and a time-series implicit representation is generated. The temporal latent feature representation is subjected to time-aware attention weighting, and dynamic weights are assigned to different time steps to obtain an enhanced temporal representation and the attention weight distribution of each time step. Regression prediction is performed based on the enhanced time-series representation to output the predicted power load values for future times; Based on the enhanced time-series representation, the attention weight distribution, and the power load forecast, a multi-level interpretability analysis is performed.
[0009] Furthermore, the multi-source heterogeneous data includes historical load data, meteorological factor data, electricity price information, and time characteristic information; The time feature information includes time periods and holiday information; the meteorological factor data includes weather features such as temperature, humidity, and wind speed. The multi-source heterogeneous data comes from different data sources and is integrated after timestamp alignment to form a historical power load dataset.
[0010] Furthermore, the preprocessing includes: The missing values in the historical power load dataset are filled using interpolation. The features of each dimension in the historical power load dataset are standardized and mapped to a distribution with a mean of zero and a variance of one. Data augmentation is performed on the standardized data, including random truncation and noise perturbation operations.
[0011] Furthermore, based on the gating mechanism, the importance of different initial input features is dynamically weighted to obtain a weighted residual feature representation, including: The initial input features are represented as a set of features with multiple dimensions: The gating value corresponding to each initial input feature is calculated using a two-layer neural network; The weighted feature is obtained by multiplying the gate value corresponding to each initial input feature by the initial input feature; The weighted features are fused with the residual features of the initial input features to obtain the weighted residual feature representation.
[0012] Furthermore, the liquid neural network adopts a neural circuit strategy structure, including a sensory layer, an interaction layer, an instruction layer, and an execution layer, specifically including: The weighted residual feature representation is used as the input to the liquid neural network:
[0013] in, This represents the weighted residual characteristic representation, where t represents the time step; This represents the d-dimensional vector space to which the feature vector belongs, where d is the feature dimension; Calculate the input signals of neurons in the sensory layer:
[0014] in, This represents the feature vector after processing by neurons in the sensory layer. For the input weight vector, This is the third bias term; Each liquid neuron in the interaction layer Its dynamic characteristics are modeled using differential equations:
[0015] in, This is the time constant of the i-th neuron, which can be dynamically adjusted according to the input; The weights represent the sparse connections between neurons in the interaction layer, where the connections can be excitatory or inhibitory. It is a non-linear activation function; Aggregate the states of all neurons in the interaction layer to form the instruction layer representation:
[0016] in, This represents the feature vector after processing by the neurons in the interaction layer. For learnable aggregate weights, It is a nonlinear function; The instruction layer output is used to obtain a timing implicit representation through the execution layer:
[0017] in, Let m represent the implicit representation of the liquid neural network at time t, where m is the number of neurons in the execution layer. These represent the dynamic representations of different features after processing by a liquid neural network.
[0018] Furthermore, time-aware attention weighting is applied to the temporal latent feature representation, and dynamic weights are assigned to different time steps to obtain an enhanced temporal representation and attention weight distribution for each time step. Specifically, this includes: The implicit temporal representation is concatenated with the temporal feature vector at the corresponding moment to obtain the concatenated time series representation; A linear mapping is performed on the concatenated time series representation based on an attention mechanism to obtain the query vector, key vector, and value vector: Attention weights are obtained by calculating and normalizing the correlation between the query vector and the key vector; the enhanced temporal representation is obtained by weighted summation of the value vector using the attention weights.
[0019] Furthermore, the aforementioned multi-level interpretability analysis includes: Based on the gradient attribution method, the global contribution of each input feature is calculated by multiplying the gradient of the predicted power load value at future time relative to the input feature with the input feature value. We perform feature attribution at the moment of mutation by tracing the sparse connection path of the liquid neural network in reverse to identify the key input features that cause the mutation in the prediction result. Based on the attention weight distribution output by the time-aware attention mechanism, the importance coefficient of each historical time step is calculated.
[0020] Furthermore, perturbation tests simulating holidays, extreme weather, or electricity price fluctuations are introduced during the training process of the liquid neural network model to verify the robustness and interpretability of the model.
[0021] Furthermore, one or more of the mean absolute error, mean square error, and mean absolute percentage error are used as evaluation metrics to assess the predictive performance of the liquid neural network model, and the model is optimized in conjunction with the results of the multi-level interpretability analysis.
[0022] On the other hand, the present invention provides a power load forecasting system based on a liquid neural network and a multi-level interpretable module, comprising: The data upload module is used to collect multi-source heterogeneous data from the power system and preprocess the multi-source heterogeneous data to construct initial input features; The feature weighting allocation module is used to dynamically weight the importance of different initial input features based on a gating mechanism to obtain a weighted residual feature representation. The time series feature extraction module is used to extract time series features from the weighted residual feature representation using a liquid neural network. It models the non-stationary dynamic characteristics of the power load through liquid neurons with adaptive time constants and sparse connections, and generates a time series implicit representation. The time-aware attention module is used to perform time-aware attention weighting on the temporal latent feature representation, assign dynamic weights to different time steps, and obtain an enhanced temporal representation and the attention weight distribution of each time step. The prediction and regression module is used to perform regression prediction based on the enhanced time-series representation and output the predicted power load value at future times. A multi-level interpretability module is used to perform multi-level interpretability analysis based on the enhanced time-series representation, the attention weight distribution, and the power load forecast. Compared with the prior art, the present invention has the following beneficial effects: In summary, by utilizing the above-mentioned technical solution of this invention, the prediction accuracy of the model is improved through modeling the non-stationary temporal characteristics of liquid neural networks; simultaneously, the interpretability of the model is enhanced by employing multiple layers of model interpretability modules, providing transparent and traceable evidence for the prediction results; furthermore, the reliability and robustness of the model under extreme operating conditions are ensured through a mutation scenario simulation and interpretation mechanism. The method and system can be directly embedded into power dispatching platforms or cloud systems, supporting real-time prediction and decision support. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the implementation of a power load forecasting method based on a liquid neural network and a multi-level interpretable module according to the present invention. Figure 2 This is a schematic diagram of the model load prediction results in a specific embodiment of the power load prediction method based on liquid neural network and multi-level interpretable module of the present invention; Figure 3 This is a schematic diagram of the interpretable results of a multi-level model for a power load prediction method based on a liquid neural network and a multi-level interpretable module according to the present invention. Figure 4This is a visualization of the time step importance analysis results of the power load forecasting method based on liquid neural network and multi-level interpretable module of the present invention. Figure 5 This is a visualization of the attribution results of abrupt time step features in the power load prediction method based on liquid neural networks and multi-level interpretable modules of this invention. Detailed Implementation
[0024] Through extensive research and practice, the inventors of this application have discovered that current electricity load forecasting methods face the following challenges: (1) Insufficient interpretability of the model: Most of the deep neural networks currently used have "black box" characteristics. Although they have improved the accuracy of load forecasting, their internal calculation and reasoning processes are difficult to present transparently and lack intuitive explanation of the forecast results. This limits the understanding and trust of power grid dispatchers in the model results and reduces their application value in actual dispatching decisions.
[0025] (2) Limited ability to model non-stationary loads: Traditional recurrent neural networks (RNN) and long short-term memory networks (LSTM) usually rely on fixed time constants and are difficult to adapt to the time-varying characteristics of load data. Therefore, they are not good at capturing dynamic fluctuations at multiple time scales and their modeling effect on non-stationary load processes is not ideal.
[0026] (3) Insufficient robustness in sudden change scenarios: In abnormal situations such as holidays, extreme weather conditions or sudden events, the power load often fluctuates violently. Existing methods show a significant increase in prediction error in such sudden change scenarios, which indicates poor robustness and makes it difficult to meet the stability requirements of power grid operation.
[0027] Based on the above considerations, this invention proposes a power load forecasting method based on liquid neural networks and multi-level interpretable modules, which mainly solves the problems of poor interpretability of current load forecasting models, limited ability to model time series features of non-stationary scenarios in practical applications, and insufficient robustness to sudden change scenarios.
[0028] To achieve the above objectives, the main concept of this invention is as follows: This paper first proposes a power load forecasting method and system based on liquid neural networks and multi-level interpretable modules. By constructing a model combining liquid neural networks and time-aware attention mechanisms, it achieves modeling of the non-stationary dynamic characteristics of power load and effective extraction of features at multiple time scales. To address the problems of insufficient model interpretability, limited ability to model non-stationary dynamics, and poor robustness under abrupt change scenarios in existing methods, this invention innovatively introduces liquid neural networks. Differential equations are used to describe the neuron states, giving them adaptive time constants and sparse connection mechanisms, thereby enhancing the ability to capture complex dynamic features of load data. By assigning differentiated weights to different times through a time-aware attention module, the contribution of key time steps to the prediction results is highlighted, improving the model's ability to characterize load fluctuations under holidays and extreme weather conditions. At the same time, a multi-level interpretability module is designed, including gradient-based feature importance analysis, time step importance assessment based on attention weights, and abrupt change feature attribution based on path tracing, providing multi-faceted interpretive basis for model prediction. This method not only improves the accuracy of power load forecasting, but also enhances the transparency and traceability of forecast results, making it easier for grid dispatchers to understand and apply. It has strong generalization ability and practical promotion value.
[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Example 1 like Figure 1 As shown, this embodiment provides a power load forecasting method based on liquid neural networks and multi-level interpretable modules. The establishment of this method includes: Step 1: Collect multi-source heterogeneous data from the power system, and preprocess the multi-source heterogeneous data to construct initial input features; Step 2: Calculate the importance of different initial input features based on a gating mechanism and dynamically weight them to obtain a weighted residual feature representation; Step 3: Use a liquid neural network to extract time-series features from the weighted residual feature representation. Model the non-stationary dynamic characteristics of the power load using liquid neurons with adaptive time constants and sparse connections, and generate a time-series implicit representation. Step 4: Perform time-aware attention weighting on the temporal latent feature representation, assign dynamic weights to different time steps, and obtain the enhanced temporal representation and the attention weight distribution of each time step; Step 5: Perform regression prediction based on the enhanced time-series representation and output the predicted power load values for future times; Step Six: Based on the enhanced time-series representation, the attention weight distribution, and the predicted power load, perform a multi-level interpretability analysis.
[0031] In step one of this embodiment, the multi-source heterogeneous data includes historical load data, meteorological factor data, electricity price information, and historical auxiliary data such as time characteristic information. Specifically, multi-source data of a power grid in a certain region of North China is selected as the research object. The data sources include historical load curves from the power grid dispatch center, meteorological indicators (temperature, humidity, wind speed, etc.) provided by the meteorological station, real-time electricity prices in the electricity market, and holiday markers. All data are timestamped to form a standardized raw dataset. Missing values were detected in the original dataset, and a combination of linear interpolation and moving average was used to repair missing points. Then, min-max normalization was applied to each feature dimension to compress all inputs into the 0-1 range. Finally, random perturbation was used to augment the data and simulate real-world data. In one implementation, step one specifically includes: Step 11: In this embodiment, the dataset includes historical load data provided by the power grid dispatch center, with a time resolution of 30 minutes, covering three consecutive years of data; meteorological data provided by the meteorological department, including temperature, relative humidity, and wind speed; electricity price data from the electricity trading market, including day-ahead and real-time electricity prices; and time feature information generated based on the calendar, including date, hour, and weekday / holiday markers.
[0032] Steps one and two involve aligning the time series data according to timestamps and storing it as a CSV data file to facilitate data management and processing, providing data support for subsequent model training. In this embodiment, data preprocessing is performed to ensure data validity. Linear interpolation is used for missing data points to ensure the continuity of the time series data, using the following formula:
[0033] in, and These are the observations adjacent to the missing point. This indicates the time corresponding to the adjacent observation before the missing point. Indicates the time corresponding to the missing point. This indicates the time corresponding to the next observation after the missing point. ; For cases where more than 3 consecutive time steps are missing, the moving average method is used:
[0034] Where k=6, it means that the average of the past 6 time steps is used as the replacement value. This indicates the data entered after the mean of missing points. These represent the observations corresponding to the past 6 time steps; Standardize all input features:
[0035] in, This represents the input features for standardization processing. For the population mean of the sample data, The population standard deviation of the sample data; Random truncation (cutting out a portion of the time window) and random noise perturbation are introduced into the training set:
[0036] in, For additive noise terms, Represented as Gaussian noise; In one implementation, step two specifically includes: Step 21: In this embodiment, the dynamic weights of different initial input features are calculated. First, a two-layer neural network is used to calculate the feature gating values:
[0037] in, The threshold value for the initial input features. This is the weight matrix. These are the first bias term and the second bias term, respectively. It is a non-linear activation function. Sigmoid function; Step 22: Multiply the gate value by the initial input features to obtain the weighted features. :
[0038] Steps two and three involve fusing the weighted features with the initial input feature residuals to obtain the weighted residual feature representation. :
[0039] In one implementation, step three specifically includes: Step 31: Extract time series features using a liquid neural network. The specific process is as follows: First, perform a linear mapping on the weighted residual feature representation in the sensory layer to obtain the sensory layer input:
[0040] in, This is the feature vector after processing by neurons in the sensory layer. For the input weight vector, This is the third bias term; Step 32, the state update of the liquid neuron is described by the following differential equation:
[0041] in, This is the time constant of the i-th neuron, which can be dynamically adjusted according to the input; The weights represent the sparse connections between neurons in the interaction layer, where the connections can be excitatory or inhibitory. It is a non-linear activation function; Step 33: Aggregate the interaction layer output into an instruction layer representation:
[0042] in, This is represented as a feature vector after processing by the neurons in the interaction layer. For learnable aggregate weights, It is a nonlinear function; Steps three and four: The execution layer generates the temporal implicit representation, the specific expression of which is as follows:
[0043] in, Let m represent the implicit representation of the liquid neural network at time t, where m is the number of neurons in the execution layer. These represent the dynamic representations of different features after processing by a liquid neural network.
[0044] In one implementation, step four specifically includes: Step 41: Input the temporal latent representation into the time-aware attention module to obtain the enhanced representation. Then, input the temporal latent representation obtained in Step 3... The input is fed into a linear transformation layer and mapped to three categories: query (Q), key (K), and value (V).
[0045] Where h represents the input temporal implicit state, These are mapping matrices, used to learn different representation spaces; Step 42: Calculate the correlation between the query vector and the key vector using the dot product operation. :
[0046] in, The dimension of the key vector; To query the correlation between vector i and key vector j, Represents the i-th query vector, This represents the j-th key vector.
[0047] The attention weights are obtained by performing Softmax normalization on the relevance scores.
[0048] weight value This indicates the importance of the i-th predicted position at the j-th historical time step; Step 43: Weight the value vectors according to the attention weights and sum them to obtain the enhanced time series representation. Through this operation, the model can automatically highlight information from special time steps such as holidays and extreme weather.
[0049] in, This represents the enhanced temporal representation. Represents the j-th value vector; In one implementation, step five specifically includes: Based on the enhanced time-series representation obtained in step four, future power load forecasts are generated. The data is input into the prediction and regression module, where it undergoes a linear transformation through a fully connected layer. The prediction results can be visualized by comparing the predicted and actual values with a line graph, making it easy to intuitively evaluate the prediction effect. Figure 2 The model's power load predictions for users at different time steps were plotted on the test set. In one implementation, step six specifically includes: Step 61: Using the gradient attribution method, calculate the sensitivity of the prediction result to each input feature:
[0050] in, Representation of features Contribution to the prediction results. Through normalization, the importance ranking of different features is obtained, such as the influence of temperature, electricity price, and historical load on the prediction. Step 62: When the prediction result undergoes a sudden change (such as a surge in load during holidays), the system tracks the sparse connection paths of the liquid neural network to pinpoint which abnormal changes in input features are most likely to cause the sudden change. For example, if the temperature rises abnormally and electricity prices fluctuate drastically during a certain period, the system can identify these features as the main driving factors of the sudden change. Step 63: Based on the attention weight distribution in Step 4, calculate the relative contribution of each historical time step in the prediction. The time step weight distribution can be displayed in the form of a heatmap to intuitively understand which historical time periods the model relies on for prediction.
[0051] Through the above multi-level interpretation process, not only can predicted values be obtained, but users can also clearly understand the feature drivers and key time points behind the predictions, thereby improving the transparency and credibility of the model results.
[0052] Furthermore, in this embodiment, perturbation tests simulating holidays, extreme weather, or electricity price fluctuations are introduced during the training process of the liquid neural network model to verify the robustness and interpretability of the model.
[0053] Furthermore, in this embodiment, one or more of the mean absolute error, mean square error, and mean absolute percentage error are used as evaluation indicators to evaluate the predictive performance of the liquid neural network model, and the model is optimized in combination with the results of the multi-level interpretability analysis.
[0054] The indicators include: Mean Absolute Error (MAE):
[0055] Mean Square Error (MSE):
[0056] Mean Absolute Percentage Error (MAPE):
[0057] The model is then optimized by incorporating explanatory metrics.
[0058] This embodiment selects a multidimensional dataset of load from a region in North China, containing seven features: date, hour, whether it is a holiday, temperature, humidity, electricity price, and power load, with a time interval of 30 minutes. The data was collected in a real environment, reflecting actual electricity consumption in daily life, enhancing the realism and practicality of the research; the dataset covers three years, facilitating time series modeling and analysis.
[0059] Figure 2 This image shows a comparison between the predicted and actual user electricity load values obtained by the liquid neural network model over multiple future time steps on the test set. As can be seen from the image, the model has high prediction accuracy in load forecasting and is able to capture the non-stationary characteristics in actual load data.
[0060] Figure 3-5The visualizations of model interpretation results at three levels—feature importance, dominant features at abrupt change times, and key time steps—show that hourly features contribute the most to the overall model and influence prediction accuracy. Meanwhile, among the dominant features at abrupt change times in random samples, temperature appears most frequently, indicating that temperature more often affects non-stationary load changes. Furthermore, the interpretation results of time step importance reveal that 11:00-13:00 and 20:00-24:00 are key time steps in the load forecasting process, which may be closely related to residents' daily lives.
[0061] Example 2 This embodiment includes a power load forecasting system based on liquid neural networks and multi-level interpretable modules. The system mainly includes the following functional modules: Data upload module: used to collect multi-source heterogeneous data from the power system, and to preprocess the multi-source heterogeneous data to construct initial input features; This module receives and stores multi-source heterogeneous data from the power system, specifically including historical load data, meteorological factors (temperature, humidity, wind speed), electricity price data, and time-related information. The module provides a unified interface supporting various data formats such as CSV, database integration, and real-time streaming uploads. Simultaneously, it performs data validity checks during the upload phase, such as timestamp alignment and outlier detection, to ensure the integrity and consistency of the input data.
[0062] The feature weighting allocation module is used to dynamically weight the importance of different initial input features based on a gating mechanism to obtain a weighted residual feature representation. The time series feature extraction module is used to extract time series features from the weighted residual feature representation using a liquid neural network. It models the non-stationary dynamic characteristics of the power load through liquid neurons with adaptive time constants and sparse connections, and generates a time series implicit representation. The time-aware attention module is used to perform time-aware attention weighting on the temporal latent feature representation, assign dynamic weights to different time steps, and obtain an enhanced temporal representation and the attention weight distribution of each time step. The prediction and regression module is used to perform regression prediction based on the enhanced time-series representation and output the predicted power load value at future times. A multi-level interpretability module is used to perform multi-level interpretability analysis based on the enhanced time-series representation, the attention weight distribution, and the power load forecast.
[0063] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0064] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations to the embodiments of the present invention fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.
[0065] All other parts not described in detail are existing technologies.
Claims
1. A power load forecasting method based on liquid neural networks and multi-level interpretable modules, characterized in that, include: Collect multi-source heterogeneous data from the power system, and preprocess the multi-source heterogeneous data to construct initial input features; The importance of different initial input features is calculated based on a gating mechanism and dynamically weighted to obtain a weighted residual feature representation. The weighted residual feature representation is extracted using a liquid neural network. The non-stationary dynamic characteristics of the power load are modeled by liquid neurons with adaptive time constants and sparse connections, and a time-series implicit representation is generated. The temporal latent feature representation is subjected to time-aware attention weighting, and dynamic weights are assigned to different time steps to obtain an enhanced temporal representation and the attention weight distribution of each time step. Regression prediction is performed based on the enhanced time-series representation to output the predicted power load values for future times; Based on the enhanced time-series representation, the attention weight distribution, and the power load forecast, a multi-level interpretability analysis is performed.
2. The power load forecasting method based on liquid neural networks and multi-level interpretable modules according to claim 1, characterized in that: The multi-source heterogeneous data includes historical load data, meteorological factor data, electricity price information, and time characteristic information; The time feature information includes time periods and holiday information; the meteorological factor data includes weather features such as temperature, humidity, and wind speed. The multi-source heterogeneous data comes from different data sources and is integrated after timestamp alignment to form a historical power load dataset.
3. The power load forecasting method based on liquid neural networks and multi-level interpretable modules according to claim 2, characterized in that: The preprocessing includes: The missing values in the historical power load dataset are filled using interpolation. The features of each dimension in the historical power load dataset are standardized and mapped to a distribution with a mean of zero and a variance of one. Data augmentation is performed on the standardized data, including random truncation and noise perturbation operations.
4. The power load forecasting method based on liquid neural networks and multi-level interpretable modules according to claim 1, characterized in that, The importance of different initial input features is calculated based on a gating mechanism and then dynamically weighted to obtain a weighted residual feature representation, including: The initial input features are represented as a set of features with multiple dimensions: The gating value corresponding to each initial input feature is calculated using a two-layer neural network; The weighted feature is obtained by multiplying the gate value corresponding to each initial input feature by the initial input feature; The weighted features are fused with the residual features of the initial input features to obtain the weighted residual feature representation.
5. The power load forecasting method based on liquid neural networks and multi-level interpretable modules according to claim 1, characterized in that, The liquid neural network adopts a neural circuit strategy structure, including a sensory layer, an interaction layer, an instruction layer, and an execution layer, specifically including: The weighted residual feature representation is used as the input to the liquid neural network: in, This represents the weighted residual characteristic representation, where t represents the time step; This represents the d-dimensional vector space to which the feature vector belongs, where d is the feature dimension; Calculate the input signals of neurons in the sensory layer: in, This represents the feature vector after processing by neurons in the sensory layer. For the input weight vector, This is the third bias term; Each liquid neuron in the interaction layer Its dynamic characteristics are modeled using differential equations: in, This is the time constant of the i-th neuron, which can be dynamically adjusted according to the input; The weights represent the sparse connections between neurons in the interaction layer, where the connections can be excitatory or inhibitory. It is a non-linear activation function; Aggregate the states of all neurons in the interaction layer to form the instruction layer representation: in, This represents the feature vector after processing by the neurons in the interaction layer. For learnable aggregate weights, It is a nonlinear function; The instruction layer output is used to obtain a timing implicit representation through the execution layer: in, Let m represent the implicit representation of the liquid neural network at time t, where m is the number of neurons in the execution layer. These represent the dynamic representations of different features after processing by a liquid neural network.
6. The power load forecasting method based on liquid neural networks and multi-level interpretable modules according to claim 1, characterized in that, The temporal latent feature representation is subjected to time-aware attention weighting, and dynamic weights are assigned to different time steps to obtain an enhanced temporal representation and attention weight distribution for each time step. Specifically, this includes: The implicit temporal representation is concatenated with the temporal feature vector at the corresponding moment to obtain the concatenated time series representation; A linear mapping is performed on the concatenated time series representation based on an attention mechanism to obtain the query vector, key vector, and value vector: Attention weights are obtained by calculating and normalizing the correlation between the query vector and the key vector; the enhanced temporal representation is obtained by weighted summation of the value vector using the attention weights.
7. A power load forecasting method based on liquid neural networks and multi-level interpretable modules as described in claim 1, characterized in that, The aforementioned multi-level interpretability analysis includes: Based on the gradient attribution method, the global contribution of each input feature is calculated by multiplying the gradient of the predicted power load value at future time relative to the input feature with the input feature value. We perform feature attribution at the moment of mutation by tracing the sparse connection path of the liquid neural network in reverse to identify the key input features that cause the mutation in the prediction result. Based on the attention weight distribution output by the time-aware attention mechanism, the importance coefficient of each historical time step is calculated.
8. The power load forecasting method based on liquid neural networks and multi-level interpretable modules as described in claim 1, characterized in that, In the training process of liquid neural network models, perturbation tests simulating holidays, extreme weather, or electricity price fluctuations are introduced to verify the robustness and interpretability of the model.
9. The power load forecasting method based on liquid neural networks and multi-level interpretable modules as described in claim 1, characterized in that... The predictive performance of the liquid neural network model is evaluated using one or more of the following metrics: mean absolute error, mean square error, and mean absolute percentage error. The model is then optimized based on the results of the multi-level interpretability analysis.
10. A power load forecasting system based on liquid neural networks and multi-level interpretable modules, characterized in that, include: The data upload module is used to collect multi-source heterogeneous data from the power system and preprocess the multi-source heterogeneous data to construct initial input features; The feature weighting allocation module is used to dynamically weight the importance of different initial input features based on a gating mechanism to obtain a weighted residual feature representation. The time series feature extraction module is used to extract time series features from the weighted residual feature representation using a liquid neural network. It models the non-stationary dynamic characteristics of the power load through liquid neurons with adaptive time constants and sparse connections, and generates a time series implicit representation. The time-aware attention module is used to perform time-aware attention weighting on the temporal latent feature representation, assign dynamic weights to different time steps, and obtain an enhanced temporal representation and the attention weight distribution of each time step. The prediction and regression module is used to perform regression prediction based on the enhanced time-series representation and output the predicted power load value at future times. A multi-level interpretability module is used to perform multi-level interpretability analysis based on the enhanced time-series representation, the attention weight distribution, and the power load forecast. The power load forecasting system based on liquid neural networks and multi-level interpretable modules is used to perform the steps in the power load forecasting method based on liquid neural networks and multi-level interpretable modules as described in any one of claims 1-9.