A small sample power consumption prediction method based on mutual information feature screening

Abstract

A small-sample electricity consumption prediction method based on mutual information feature selection is proposed. This method utilizes historical electricity consumption sequences and multivariate feature data of the target electricity consumer. It constructs time lag sequences with different lag orders for the multivariate features to form a candidate feature set. Under limited sample conditions, multiple subsequences of electricity consumption are sampled, and the mutual information value between each time lag feature in the candidate feature set and the sampled electricity consumption subsequence is calculated and analyzed. A comprehensive score is constructed based on the mean and stability of the mutual information to determine the optimal number of features and select a subset of strongly correlated features. The strongly correlated feature subset and historical electricity consumption are input together into a time series prediction model for training or prediction, outputting the predicted electricity consumption result at the target time. This method can stably select a small number of key features in small-sample scenarios, reducing sample fluctuation interference and improving prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of power data analysis and intelligent forecasting technology, and in particular to a small-sample power consumption forecasting method based on mutual information feature screening, which can be used for small-sample power consumption forecasting. Background Technology

[0002] Electricity consumption forecasting is a crucial foundation for power system dispatching and operation, demand response, energy management, and energy conservation optimization. Existing forecasting methods typically rely on historical electricity consumption time series data and incorporate external influencing factors such as environmental factors, equipment operating status, time factors, and user behavior to construct forecasting models.

[0003] In recent years, deep learning models such as Long Short-Term Memory (LSTM) networks have achieved good results in electricity consumption prediction tasks due to their ability to model nonlinear time series. However, in practical applications, electricity consumption prediction often suffers from the problem of insufficient historical data and small sample sizes. Furthermore, electricity consumption prediction typically involves multiple features with high dimensionality and often contains redundant or noisy features. Directly inputting these features into the model would further increase model complexity and reduce prediction accuracy and generalization ability.

[0004] To address the aforementioned issues, this invention proposes a small-sample electricity consumption prediction method based on mutual information feature screening. This method can effectively achieve stable feature screening under small sample conditions, reduce sample fluctuation interference, and accurately characterize the nonlinear relationship between external features and electricity consumption. Summary of the Invention

[0005] This invention aims to solve the technical problems of existing electricity consumption prediction methods in small sample scenarios, where key influencing features are difficult to identify stably, and the redundancy and noise in multi-dimensional features lead to a decrease in prediction accuracy and insufficient generalization ability after direct modeling.

[0006] To address the aforementioned technical problems, this invention provides a small-sample electricity consumption prediction method based on mutual information feature screening, the method comprising the following steps:

[0007] First, using a small sample of electricity consumption data and multivariate feature data related to the electricity user, including environmental features, equipment operating status features, time features, and user behavior features, a time-lag feature sequence is constructed from the multivariate feature data to form a candidate feature set. The candidate feature set and the electricity consumption data are then preprocessed.

[0008] Then, multiple subsequence samplings are performed on the small sample of electricity consumption data. The mutual information value is calculated between each time lag feature in the candidate feature set and the sampled electricity consumption subsequence. The mutual information stability statistical analysis is then performed on the calculation results.

[0009] Finally, by analyzing the results, a subset of features strongly correlated with electricity consumption is selected. This strongly correlated subset of features and the electricity consumption data are then input into an LSTM for prediction, and the predicted electricity consumption results are output.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) By combining multiple subsequence sampling with mutual information analysis, key features affecting changes in electricity consumption can be identified more effectively under small sample conditions; (2) By introducing mutual information stability analysis, the impact of single sampling fluctuations on feature evaluation results can be reduced, and the stability of feature selection can be improved; (3) By selecting strongly correlated feature subsets, redundant and noisy feature interference can be reduced, and the model complexity can be reduced; (4) Inputting the selected features into the LSTM model for prediction can improve the accuracy of electricity consumption prediction in small sample scenarios. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a small-sample electricity consumption prediction method based on mutual information feature screening according to the present invention.

[0012] Figure 2 This is a graph showing the prediction results of an example of the present invention. Detailed Implementation

[0013] A small-sample electricity consumption prediction method based on mutual information feature screening, the method comprising the following operations:

[0014] First, use a small sample of electricity consumption data. and multi-dimensional characteristic data related to the electricity user. ,in , For small sample data length, This refers to the number of multi-dimensional features. Multi-dimensional features include: environmental features, equipment operating status features, time features, and user behavior features.

[0015] A time-lag feature sequence is constructed from the multivariate feature data to form a candidate feature set. Let the maximum lag order be... , for the Constructing lagged features from individual features:

[0016]

[0017] in Indicates the first Each feature is lagging The hysteresis characteristics after the order, At this point, the candidate feature set is formed. for:

[0018]

[0019] At this point, the total number of candidate features is .

[0020] Preprocessing of candidate feature set and electricity consumption data:

[0021] Standardization of electricity consumption sequences:

[0022]

[0023] in, This represents the standardized electricity consumption data. This represents the average electricity consumption. This represents the standard deviation of electricity consumption;

[0024] Standardize the candidate features:

[0025]

[0026] in, This represents the standardized candidate feature data. Representing candidate features The mean, Representing candidate features The standard deviation of , at this point, the candidate feature set is represented as:

[0027]

[0028] in, This represents the standardized set of candidate features.

[0029] Secondly, to improve the robustness of mutual information under small sample conditions, the electricity consumption is sampled multiple times from a subsequence. Let the number of sampling times be... Each sampling yields a length of subsequence of:

[0030]

[0031] in, This represents the subsequence obtained after sampling. , Indicates the first The starting position of the next sampling The sampling method is: random block sampling;

[0032] Correspondingly, for each candidate feature Extract the feature subsequence aligned with the electricity consumption subsequence. :

[0033]

[0034] For each sample, calculate the electricity consumption of the subsequence. With the corresponding candidate feature subsequence Mutual information values ​​between :

[0035]

[0036] Mutual information is defined as:

[0037]

[0038] in, and They are two random variables. express Values and Values The joint probability, express marginal distribution, express The marginal distribution.

[0039] Finally, the mutual information results are analyzed to select a subset of features strongly correlated with electricity consumption. Based on the mutual information values, this invention introduces a mutual information stability index, using the coefficient of variation. The relative dispersion of the mutual information calculation results is characterized. Mutual information mean. for:

[0040]

[0041] coefficient of variation for:

[0042]

[0043] in To prevent extremely small constants with a denominator of zero, It is the standard deviation of mutual information;

[0044] A smaller coefficient of variation indicates less change in mutual information value with sampling, and higher stability. To simultaneously consider the magnitude and stability of mutual information, this invention constructs a comprehensive evaluation index. :

[0045]

[0046] This comprehensive evaluation index is formally equivalent to imposing a stability penalty on the mean of mutual information; when the mean of mutual information... The larger the coefficient of variation, the higher the coefficient of variation. The smaller the value, the higher the stability of mutual information; at this point, the comprehensive evaluation index... The higher;

[0047] Select the top K candidate features with the highest scores as the strongly correlated feature subset. :

[0048]

[0049] Using historical electricity consumption sequences as the main sequence input branch, and strongly correlated feature subsets... As a multi-feature input branch, a multi-input prediction model is constructed. A dual-branch LSTM structure is used to extract main sequence features and multi-features separately, which are then fused through a fusion layer, and finally output as predicted values ​​through a fully connected layer. Then, the obtained predicted values ​​are de-standardized to obtain the predicted values ​​at the true scale. :

[0050]

[0051] To verify the effectiveness of the method of the present invention, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are selected for verification in the examples of the present invention. The formulas for the two are as follows:

[0052]

[0053]

[0054] in, It is a test set Predicted electricity consumption at any given time. It is a test set The actual power consumption at any given moment. This is the predicted future time step for the test set.

[0055] This invention uses data from a certain province, including electricity consumption, highest temperature, maximum load, daily average humidity, population migration index, and holiday data. The prediction result curve is shown below. Figure 2 As shown, in the optimal case ( The method of this invention achieves significant improvements compared to direct prediction by LSTM: RMSE is reduced from 181.1722 to 172.0617, an improvement of approximately 5.03%; MAPE is reduced from 2.22% to 2.07%, an improvement of approximately 6.81%.

[0056] Table 1. Model Evaluation Index Results

[0057] Model RMSE MAPE LSTM direct prediction 181.1722 2.22% This invention predicts 172.0617 2.07%

[0058] As can be seen from the results in Table 1, the predicted effect of the present invention is better than the direct prediction effect, which proves the advantages of the present invention.

Claims

1. A small sample power consumption prediction method based on mutual information feature screening, characterized in that, The method includes the following steps: First, using a small sample of electricity consumption data and multivariate feature data related to the electricity consumer, a time-lag feature sequence is constructed from the multivariate feature data to form a candidate feature set. The candidate feature set and the electricity consumption data are then preprocessed. Second, the electricity consumption data from the small sample is sampled multiple times, and the mutual information value between each time-lag feature in the candidate feature set and the sampled electricity consumption subsequence is calculated. Let the number of samplings be... Each sampling yields a length of subsequence of: in, This represents the standardized electricity consumption data. This represents the subsequence obtained after sampling. , Indicates the first The starting position of the next sampling The sampling method is random block sampling, corresponding to each candidate feature. Extract the feature subsequence aligned with the electricity consumption subsequence. : in, Indicates the first Each feature is lagging The hysteresis characteristics after the order, , express Standardized candidate feature data; For each sample, calculate the electricity consumption of the subsequence. With the corresponding candidate feature subsequence Mutual information values ​​between : ； Finally, the mutual information results are analyzed, and the feature subsets strongly related to the power consumption are selected by analyzing the results The relative dispersion degree of the mutual information calculation results is described, and the mutual information mean is: Coefficient of variation is: Wherein To prevent the denominator of the minimum constant, Is the standard deviation of mutual information, the smaller the coefficient of variation, the smaller the change of mutual information value with sampling, the higher the stability, the comprehensive evaluation index is constructed by the present application : This comprehensive evaluation index is formally equivalent to applying a stability penalty to the mean mutual information of candidate features; when the mean mutual information... The larger the coefficient of variation, the higher the coefficient of variation. The smaller the value, the higher the stability of mutual information; at this point, the comprehensive evaluation index... The higher the score, the higher the K candidate features are selected as the strongly correlated feature subset. : ， The characteristic subset is input into the LSTM together with the power consumption data to perform prediction, and a prediction result of the power consumption is output. 2.The small sample power consumption prediction method based on mutual information feature screening according to claim 1, wherein, The multivariate feature data and electricity consumption data are preprocessed. First, a time lag feature sequence is constructed from the multivariate feature data to form a candidate feature set. The maximum lag order is set as follows: , for the Constructing lagged features from individual features: At this time, the candidate feature set formed is: The total number of selected features at this time is The power consumption and the candidate feature set are respectively standardized, and the power consumption sequence is standardized: wherein, denotes the mean of the power consumption, denotes the standard deviation of the power consumption, normalizing the candidate features: wherein, denotes the mean of the candidate features , denotes the standard deviation of the candidate features , and wherein the set of candidate features is denoted as: wherein, denotes the normalized candidate feature set.

Images