Deep learning-based short-term power load prediction method

Abstract

The invention discloses a short-term power load prediction method based on deep learning, and the method comprises the steps: S1, collecting historical power load data and corresponding weather and holiday and festival information, S2, carrying out the preprocessing of the historical data based on the historical power load data and the change of a power utilization behavior mode of a user, reducing the data inconsistency, and carrying out the prediction of a short-term power load. The method comprises the following steps: S1, processing historical power load data, weather information and holiday and festival information, S3, constructing a deep learning model which inputs processed historical power load data, weather information and holiday and festival information, S4, training the constructed deep learning model by using a training data set and optimizing model parameters, and S5, carrying out short-term power load prediction by using the trained deep learning model and generating a prediction result. According to the short-term power load prediction method based on deep learning, the problem of inconsistency of historical data caused by frequent change of power utilization habits of users in a changeable user behavior mode in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting and user behavior analysis, and in particular to a short-term power load forecasting method based on deep learning. Background Art

[0002] In modern power systems, the analysis of user electricity consumption behavior patterns is an important part of optimizing power dispatch and improving the efficiency of power resource utilization. By analyzing the historical data of user electricity consumption habits, users' electricity demand can be predicted to provide support for the stable operation of the power system. However, user behavior patterns are often affected by many factors, such as seasonal changes, emergencies, or electricity price adjustments, which leads to frequent changes in users' electricity consumption habits.

[0003] In the prior art, analysis based on user electricity usage habits mainly relies on the integration and modeling of historical data. However, due to the variability of user electricity usage behavior patterns, historical data is prone to inconsistency problems. This inconsistency is manifested in the poor timeliness and representativeness of the data, which makes it difficult to accurately reflect the user's current or future electricity usage behavior. For example, there are certain outliers or highly volatile intervals in the user's electricity usage data. If these data are not effectively identified and processed, they will have an adverse effect on the accuracy of the prediction model. In addition, traditional electricity usage behavior modeling methods often lack sensitivity to dynamic changes. Most methods are based on static assumptions for modeling and cannot effectively capture the dynamic change characteristics in user behavior patterns. This not only reduces the accuracy of user electricity usage behavior pattern analysis, but also leads to insufficient responsiveness of the power system in balancing supply and demand, affecting power supply reliability and economy. Summary of the invention

[0004] The purpose of the present invention is to provide a short-term power load forecasting method based on deep learning to solve the problem of inconsistency in historical data caused by frequent changes in user electricity consumption habits under changeable user behavior patterns in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: a short-term power load forecasting method based on deep learning, the method comprising:

[0006] S1. Collect historical power load data and corresponding weather and holiday information;

[0007] S2. Based on the historical power load data and the changes in the user's power consumption behavior pattern, the historical data is preprocessed to reduce data inconsistency. Specifically, the historical load data is smoothed, abnormal fluctuations and peak data are removed, and the power consumption data is seasonally decomposed to extract long-term trends, cyclical changes and random components. Data from different sources are aligned by timestamps and missing values ​​are filled. The formula for smoothing the historical load data is:

[0008] S t =αY t +(1-α)S t-1 ;

[0009] Among them, S t It represents the historical power load data value at a certain moment after smoothing, Y t represents the original power load data actually measured at a certain moment, α represents the smoothing factor, S t-1 It represents the result of smoothing the power load data at the previous moment, and t represents time;

[0010] S3, building a deep learning model, the model input is processed historical power load data, weather and holiday information;

[0011] S4. Use the training data set to train the constructed deep learning model and optimize the model parameters;

[0012] S5. Apply the trained deep learning model to perform short-term power load forecasting and generate forecast results.

[0013] Preferably, the S1 comprises:

[0014] The geographic information of the power load collection points is spatially matched with the meteorological data, and the sampling point data is spatially interpolated to estimate the meteorological conditions at other locations and generate a continuous spatiotemporal meteorological data table. Combined with the calendar data of holidays and weekdays, the load data and meteorological data are time-annotated to form a unified spatiotemporal sequence input. The calculation formula for spatial interpolation of the sampling point data is:

[0015] Z(s)=μ(s)+∈(s);

[0016] Among them, Z(s) represents a target value at the geographic space location s, s represents the geographic coordinate location of the power load collection point or meteorological station, μ(s) represents the overall trend or long-term stable value of data in a large range, and ∈(s) represents the random error term.

[0017] Preferably, S3 includes:

[0018] Construct a one-dimensional convolutional network, perform convolution operations on time series data, extract local features, and combine the bidirectional long short-term memory network to capture the bidirectional dependency of the time series. The one-dimensional convolution operation formula is:

[0019]

[0020] Among them, y[i] represents the eigenvalue or the i-th element in the feature sequence obtained by the one-dimensional convolution operation, x[i] represents the i-th data point in the original data sequence input to the deep learning model, w[k] represents the weight coefficient of the convolution kernel for the k-th position in the input sequence during the convolution operation, k represents the position index in the convolution kernel, K represents the size of the convolution kernel, and i represents the starting position index of the input sequence.

[0021] Preferably, S4 includes:

[0022] The historical load data is divided into multiple training tasks by time period, the changes in different power consumption scenarios are simulated, and the initial parameters of the deep learning model are adjusted. The specific formula is:

[0023]

[0024] Among them, θ * represents the final parameter value of the deep learning model after meta-learning optimization, θ represents the current parameter value of the deep learning model, and α represents the learning rate. represents the gradient operator, Represents the loss function for evaluating the overall performance of the model on different tasks.

[0025] Preferably, S5 includes:

[0026] Use the trained deep learning model to generate load forecast results, and use known observation data to calibrate the posterior probability of the forecast results. The specific formula is:

[0027]

[0028] Among them, P(X|E) represents the probability value of predicting short-term power load X under the condition of known observation data E, P(E|X) represents the probability of observation data E appearing under a certain power load X, P(X) represents the initial probability of power load X in the absence of observation data E, and P(E) represents the overall probability of observation data E appearing.

[0029] Preferably, S2 further includes:

[0030] Identify different electricity consumption behavior patterns of users through cluster analysis;

[0031] Based on the time series analysis of the user's electricity consumption behavior pattern, the change points of the electricity consumption behavior pattern are detected;

[0032] Identify and remove outliers from historical data;

[0033] Based on the detected change points in electricity usage behavior patterns, historical data is processed in segments.

[0034] Preferably, the time series analysis based on the user's electricity consumption behavior pattern to detect the change point of the electricity consumption behavior pattern includes:

[0035] Collect hourly power load data and corresponding meteorological information;

[0036] Calculate the rate of change of the electricity consumption behavior pattern in each time period ΔC = (P(t) - P(t1)) / P(t1), where P(t) represents the electricity consumption at time t, and P(t1) represents the electricity consumption at time t1;

[0037] Based on the rate of change ΔC, if ΔC>T h , it is considered that a change point in the electricity consumption behavior pattern has occurred in this time period, where T h is a pre-set threshold;

[0038] Record the time of change points and electricity usage behavior patterns.

[0039] Preferably, the calculation of the change rate of the electricity consumption behavior pattern in each time period ΔC=(P(t)-P(t1)) / P(t1) includes:

[0040] Extract the power consumption data P(t) and P(t1) at time t and time t1;

[0041] Calculate the power consumption change rate ΔC = (P(t) - P(t1)) / P(t1), where P(t) represents the power consumption at time t, and P(t1) represents the power consumption at time t1;

[0042] Based on the calculation result of ΔC, if ΔC>T h , then record the change point;

[0043] T h The value range of is [0.05, 0.2], which is used to distinguish significant change points from small fluctuations.

[0044] Preferably, the calculation result based on ΔC is, if ΔC>T h , then the record change points include:

[0045] Compare ΔC with the preset threshold T h , if ΔC>T h ,Record time t as the change point, record the change type of the power consumption behavior pattern based on the change point, and update the statistical parameters of the power consumption behavior pattern after the change point.

[0046] Preferably, the change type of the electricity consumption behavior pattern recorded based on the change point includes:

[0047] Identify the types of electricity consumption behavior patterns before and after the change, classify the change types of electricity consumption behavior patterns based on the change points and pattern matching results, and generate a pattern change type report.

[0048] It can be seen from the above technical solution that the present invention has the following beneficial effects:

[0049] This short-term power load forecasting method based on deep learning collects historical power load data and corresponding meteorological and holiday information, and based on the changes in historical power load data and user electricity consumption behavior patterns, pre-processes the historical data to reduce data inconsistency and constructs a deep learning model. The model input is the processed historical power load data, meteorological and holiday information. The constructed deep learning model is trained using a training data set, the model parameters are optimized, and the trained deep learning model is used to perform short-term power load forecasting to generate forecast results. This reduces the inconsistency of historical data, significantly improves the timeliness and representativeness of the data, provides high-quality data input for the forecasting model, and overcomes the limitations of traditional static modeling methods. By dynamically adjusting model parameters and responding to changes in user behavior patterns in real time, the prediction is made more accurate and reliable, and a more precise short-term prediction of users' future electricity demand can be made, thereby effectively supporting the dispatch optimization and stable operation of the power system, improving the responsiveness of the power system in balancing supply and demand, reducing the problem of over or under power supply due to prediction errors, improving the reliability and economy of power supply, and being able to quickly adapt to the load changes in different power consumption scenarios. It provides support for the efficient use of power resources and scientific dispatch, and solves the problem of inconsistency in historical data caused by frequent changes in user electricity usage habits under changeable user behavior patterns in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, the present invention provides a technical solution: a short-term power load forecasting method based on deep learning, the method comprising:

[0053] S1. Collect historical power load data and corresponding weather and holiday information;

[0054] S2. Based on the historical power load data and the changes in the user's power consumption behavior pattern, the historical data is preprocessed to reduce data inconsistency. Specifically, the historical load data is smoothed, abnormal fluctuations and peak data are removed, and the power consumption data is seasonally decomposed to extract long-term trends, cyclical changes and random components. Data from different sources are aligned by timestamps and missing values ​​are filled. The formula for smoothing the historical load data is:

[0055] St=αY t +(1-α)S t-1 ;

[0056] Among them, S t It represents the historical power load data value at a certain moment after smoothing, Y t represents the original power load data actually measured at a certain moment, α represents the smoothing factor, S t-1 It represents the result of smoothing the power load data at the previous moment, and t represents time;

[0057] S3, building a deep learning model, the model input is processed historical power load data, weather and holiday information;

[0058] S4. Use the training data set to train the constructed deep learning model and optimize the model parameters;

[0059] S5. Apply the trained deep learning model to perform short-term power load forecasting and generate forecast results.

[0060] In the above method, first, comprehensive historical power load data and its associated meteorological conditions (such as temperature, humidity, wind speed, etc.) and external factors such as holidays are collected through S1. Power load has strong time dependence and nonlinear characteristics, and meteorological conditions and holiday information can significantly affect the user's power consumption pattern. Then, the historical load data is preprocessed in multiple steps through S2, and the data is denoised using a smoothing formula to remove sudden outliers to ensure the continuity and consistency of the data. At the same time, the power load is divided into long-term trends, periodic changes, and random components using decomposition technology, which can better identify and adapt to features on different time scales. This process combines data alignment and missing value filling to ensure the integrity and accuracy of the input data. Subsequently, in step S3, a deep learning model is constructed using historical data. The model architecture can be a long short-term memory network, a convolutional neural network, or a hybrid model, and the input features include processed historical load data and external features (meteorological and holiday information). The model is trained through S4, and the parameters of the model are continuously optimized so that it can accurately learn the laws of historical data and predict future loads. In the final step S5, the trained model is used to predict the power load situation in the short term, and the results can directly provide a basis for power dispatching and energy distribution.

[0061] By preprocessing historical data, the interference of outliers in the data is significantly reduced, and the robustness and prediction accuracy of the model are improved. The method comprehensively considers external factors such as electricity consumption behavior patterns, meteorological conditions and holidays, so that the prediction results are more in line with the actual scenario needs. The deep learning technology is used to fully explore the nonlinear laws in the power load data and improve the modeling ability of complex time series data. The method has a clear structure and is easy to implement. It can be widely used in actual scenarios such as power dispatching, load optimization and energy distribution, and has a high promotion value.

[0062] S1 includes spatial matching of the geographic information of the power load collection points with the meteorological data, spatial interpolation of the sampling point data, estimation of the meteorological conditions at other locations and generation of a continuous spatiotemporal meteorological data table, combined with the calendar data of holidays and weekdays, time marking of the load data and meteorological data to form a unified spatiotemporal sequence input, and the calculation formula for spatial interpolation of the sampling point data is:

[0063] Z(s)=μ(s)+∈(s);

[0064] Among them, Z(s) represents a target value at the geographic space location s, s represents the geographic coordinate location of the power load collection point or meteorological station, μ(s) represents the overall trend or long-term stable value of data in a large range, and ∈(s) represents the random error term.

[0065] In the above method, in order to solve the discontinuity and local missing problems in the spatial distribution of power load data and meteorological data, the geographical information of the power load collection points is spatially matched with the meteorological data to establish unified characteristics related to time and spatial position. Specifically, the geographical coordinate position of the power load collection point and its corresponding meteorological site information are first obtained. On this basis, the spatial interpolation algorithm is used to expand the sampling point data to estimate the meteorological conditions in the area where the collection points are not deployed. The interpolation formula is: Z(s) = μ(s) + ∈(s);

[0066] The μ(s) in the formula is obtained by analyzing the long-term data trend and overall pattern of the entire region, and is used to reflect the stable part of the power load, while the random error term ∈(s) is used to characterize short-term fluctuations and uncertainties. Finally, the generated interpolation data and the actual collected data are unified in a spatiotemporal meteorological data table, combined with holiday and weekday information to form a complete input feature sequence, ensuring that the subsequent deep learning model can make full use of time and space characteristics for load forecasting.

[0067] The spatial interpolation method is used to fill the geographical gaps in the data collected at the collection points, making the spatial coverage of load data and meteorological data more complete and reducing the impact of insufficient data on the prediction results. The data are integrated into a unified spatiotemporal feature sequence by combining geographic information, time tags and holiday effects, ensuring that the deep learning model can fully capture the spatiotemporal dynamic changes in the power load. The introduction of the interpolation algorithm significantly reduces data deviation, allowing the model to more accurately reflect the actual load and meteorological conditions in the region, thereby improving the accuracy of load forecasting. This method is suitable for different geographical regions and meteorological conditions, especially for scenarios with sparse data collection points, and has strong adaptability and promotion value.

[0068] S3 includes building a one-dimensional convolutional network, performing convolution operations on time series data, extracting local features, and combining a bidirectional long short-term memory network to capture the bidirectional dependency of the time series. The one-dimensional convolution operation formula is:

[0069] Among them, y[i] represents the eigenvalue or the i-th element in the feature sequence obtained by the one-dimensional convolution operation, x[i] represents the i-th data point in the original data sequence input to the deep learning model, w[k] represents the weight coefficient of the convolution kernel for the k-th position in the input sequence during the convolution operation, k represents the position index in the convolution kernel, K represents the size of the convolution kernel, and i represents the starting position index of the input sequence.

[0070] Through one-dimensional convolution operation, the model can extract local features in time series data. The convolution kernel moves on the time series in the form of a sliding window, using the formula The local area of ​​the input sequence is weighted and summed to extract the data features near each location. This method can effectively extract the short-term change pattern in the time series, such as the peak characteristics of peak power consumption and the drastic changes in meteorological conditions. In order to capture the long-term dependencies in the power load time series and the reverse impact of future information, a bidirectional long short-term memory network is further adopted, which can learn in both the front and back directions of the time series to capture the potential complex dependencies in the load data, such as how the meteorological changes in the past few days affect today's power consumption, or the recovery pattern of power consumption after holidays. While using a one-dimensional convolutional network to extract short-term features, the bidirectional long short-term memory network is combined to capture long-term dependencies, which enhances the model's ability to express multi-time scale features. The capture of two-way information enables the model to consider past and future bidirectional time dependencies, significantly improving the accuracy of load forecasting. Compared with the traditional fully connected network, the one-dimensional convolution operation has lower computational complexity and good parallelism, which improves the training and reasoning speed of the overall model. The combined model of convolutional network and Bi-LSTM is stable in different scenarios and data sets, more adaptable, and can adapt to complex and changing power consumption patterns.

[0071] S4 includes dividing the historical load data into multiple training tasks by time period, simulating the changes in different power consumption scenarios, and adjusting the initial parameters of the deep learning model. The specific formula is:

[0072]

[0073] Among them, θ * represents the final parameter value of the deep learning model after meta-learning optimization, θ * represents the current parameter value of the deep learning model, α represents the learning rate, represents the gradient operator, Represents the loss function for evaluating the overall performance of the model on different tasks.

[0074] The historical load data is divided into multiple training tasks according to time periods (such as seasons, months, working days and holidays, etc.). These tasks correspond to different power consumption scenarios (such as peak power consumption period, off-peak power consumption period, load conditions when weather conditions change drastically, etc.), so as to simulate various actual application scenarios that the model may face. The optimization method of meta-learning is adopted, and the formula

[0075] Update the model parameters θ. In each iteration, calculate the overall loss of the model on multiple tasks Adjust the parameters through the gradient descent algorithm and gradually find the initial parameter value θ that performs well on all tasks * The optimized parameter θ *It has higher generalization ability and can quickly converge and achieve high prediction accuracy even in new and unseen tasks (such as changes in electricity demand patterns). Through this optimization process, the deep learning model can fully learn the diversity of historical load data during the training phase and provide high-quality model initialization parameters for subsequent short-term load forecasting. By dividing and simulating historical data, the model can adapt to a variety of changes in electricity consumption scenarios, such as seasonal peak loads, load fluctuations caused by sudden events, etc. The parameters optimized by meta-learning can adapt to new scenarios or tasks more quickly, improving the prediction accuracy of the model in different scenarios. The optimized initial parameters enable the model to converge quickly in the subsequent training process, reducing the time cost of model debugging and training. Considering multiple tasks, the optimized loss function can balance the performance of the model on different tasks and avoid overfitting the model to a single scenario.

[0076] S5 includes using the trained deep learning model to generate load forecast results and using known observation data to calibrate the posterior probability of the forecast results. The specific formula is:

[0077]

[0078] Among them, P(X|E) represents the probability value of predicting short-term power load X under the condition of known observation data E, P(E|X) represents the probability of observation data E appearing under a certain power load X, P(X) represents the initial probability of power load X in the absence of observation data E, and P(E) represents the overall probability of observation data E appearing.

[0079] Through the trained deep learning model, based on the processed input data (including historical power load data, meteorological conditions, holidays, etc.), a short-term power load forecast result X is generated. This result provides the model with a preliminary estimate of power demand in the future time period. Use the currently known observation data E, such as actual meteorological observations (temperature, humidity, etc.), the latest holiday adjustment information or other real-time data, as a reference for correcting the forecast results. These observation data can reflect the latest changes in power demand in the short term. Through the formula

[0080] Correct the prediction results. The Bayesian calibration process can comprehensively consider the initial prediction value P(X), the conditional probability P(E|X) of the observed data E, and the overall probability P(E) of the observed data, so as to adjust the prediction results to make them more consistent with the current actual situation. Finally, the calibrated prediction result P(X|E) is calculated based on the posterior probability. The calibrated prediction results are more accurate and reliable and can be directly used in application scenarios such as power grid scheduling and load optimization. By using the Bayesian calibration mechanism and dynamically adjusting the model prediction results in combination with the observed data, the accuracy and credibility of the prediction are significantly improved. By introducing the latest observed data, the prediction results can be updated in real time to better adapt to the rapid changes in short-term power load. The posterior probability calibration method can effectively deal with the uncertainty in the prediction results and reduce the errors of model prediction under abnormal conditions. This method can flexibly adjust the calibration process according to different types of observation data (such as meteorological, load or socio-economic indicators), and is suitable for power load prediction under various scenarios and conditions.

[0081] S2 also includes identifying different electricity consumption behavior patterns of users through cluster analysis; detecting the change points of electricity consumption behavior patterns based on the time series analysis of the electricity consumption behavior patterns of users; identifying and eliminating outliers in historical data; and processing historical data in segments based on the detected change points of electricity consumption behavior patterns. Clustering the historical load data of users is performed using a clustering analysis algorithm, and users are divided into multiple electricity consumption behavior pattern groups according to the time series characteristics of the load data. For example, electricity consumption behavior can be divided into categories such as "peak electricity consumption type", "seasonal fluctuation type" or "random fluctuation type". In this way, the electricity consumption patterns of different users or scenarios can be identified, providing a basis for subsequent data processing and analysis. Time series analysis is performed on the identified electricity consumption behavior patterns to detect the change points of electricity consumption behavior. Change point detection can be based on a mutation point detection algorithm to determine whether the electricity consumption pattern has changed significantly (such as the change of electricity consumption pattern after the implementation of energy-saving measures). By locating these change points, historical data can be divided into several segments, which is convenient for modeling and analysis of each segment of data separately. For outliers in historical data, statistical methods (such as box plot method or 3σ principle) or anomaly detection algorithms based on machine learning are used to identify and eliminate outliers. Outliers may come from equipment failures, collection errors or emergencies. Eliminating these data can significantly improve the cleanliness of historical data and the robustness of the model. According to the detected change points of power consumption behavior patterns, the historical data is divided into several time periods, each of which corresponds to a relatively stable power consumption pattern. The segmented data is subjected to feature extraction, trend analysis and standardization respectively to ensure that the features of each segment of data are more consistent, thereby improving the model's adaptability to data and prediction accuracy. Through cluster analysis and segmentation processing, complex historical data are decomposed into multiple relatively stable electricity consumption patterns, which improves data consistency and model training effect. The anomaly detection method is used to clean up noise points and erroneous records in historical data, significantly improve data quality, and reduce the error rate of the prediction model. The detection of change points in electricity consumption behavior patterns enables the model to identify changes in users' electricity consumption habits, which helps to improve the adaptability to nonlinear and sudden changes in data. The data after segmentation processing is easier to carry out targeted modeling, providing high-quality input features for the model and improving the accuracy and stability of the prediction. This method can adapt to the diverse electricity consumption behaviors of different types of users and has high practical value for household users, industrial users, and commercial users.

[0082] Based on the time series analysis of the user's electricity consumption behavior pattern, the change point of the electricity consumption behavior pattern is detected, including collecting hourly power load data and corresponding meteorological information; calculating the change rate of the electricity consumption behavior pattern in each time period ΔC = (P(t) - P(t1)) / P(t1), where P(t) represents the electricity consumption at time t, and P(t1) represents the electricity consumption at time t1; based on the change rate ΔC, if ΔC>T h, it is considered that a change point in the electricity consumption behavior pattern has occurred in this time period, where T h is a pre-set threshold; records the time of the change point and the power consumption behavior pattern.

[0083] Collect power load data and meteorological conditions (such as temperature, humidity, wind speed, etc.) on an hourly basis to ensure coverage of typical time periods and meteorological change scenarios throughout the year. This information provides a complete data set that relates power consumption behavior to environmental conditions. The rate of change of power consumption between adjacent time periods is calculated using the formula ΔC = (P(t) - P(t1)) / P(t1). The rate of change ΔC reflects the relative intensity of power consumption changes and can quantify the changing characteristics of power consumption behavior patterns in a short period of time. Compare the calculated ΔC with the threshold T h If ΔC exceeds T h , it is determined that a significant change in the electricity consumption behavior pattern has occurred during this time period. For example, the arrival of hot weather or holidays may cause a drastic change in electricity consumption. Threshold T h It can be determined based on the statistical results of historical data, for example, it can be set as the 95% quantile of the change rate distribution to ensure that the detected change points are statistically significant. The occurrence time of each change point and the corresponding power consumption behavior pattern are recorded, and the time distribution diagram of the power consumption pattern change is drawn in combination with historical data, providing a basis for segmented processing and prediction model optimization. Through the calculation of the power consumption change rate and threshold determination, the significant change points of power consumption behavior can be accurately identified, providing a quantitative basis for the analysis of power consumption behavior patterns. The method can quickly locate abnormal power consumption behaviors (such as sudden increases or decreases in load), which is helpful for the implementation of real-time dispatch and early warning mechanisms of power systems. The recording of change point information provides support for subsequent data segmentation processing, so that the model can better adapt to different power consumption behavior patterns and improve prediction accuracy. The method can be applied to a variety of power consumption scenarios (such as residential, commercial or industrial users), with high versatility and practical application value.

[0084] Calculating the change rate of the power consumption behavior pattern in each time period ΔC=(P(t)-P(t1)) / P(t1) includes extracting the power consumption data P(t) and P(t1) at time t and time t1; calculating the power consumption change rate ΔC=(P(t)-P(t1)) / P(t1), where P(t) represents the power consumption at time t, and P(t1) represents the power consumption at time t1; based on the calculation result of ΔC, if ΔC>T h , then record the change point; T h The value range of is [0.05, 0.2], which is used to distinguish significant change points from small fluctuations.

[0085] Hourly power load data P(t) and power consumption data P(t1) of the previous moment are extracted from historical data. These data points usually come from smart meters or power monitoring systems, which can reflect the dynamic changes of power load in a short period of time. Based on the extracted data, the power consumption change rate ΔC in each time period is calculated using the formula ΔC = (P(t) - P(t1)) / P(t1). This formula quantifies the change in power consumption behavior pattern by the relative degree of change in power consumption. Compare the calculated ΔC with the preset threshold Th. When ΔC exceeds Th, it is considered that the current time period has changed significantly, and the time of the change point and related information are recorded. The recorded information includes the specific time of the change point, the magnitude of the change in power consumption, and possible associated environmental conditions (such as weather or holidays). Change rate threshold T h The range of is set to [0.05, 0.2], where T h =0.05 is used to capture slight changes and is suitable for highly sensitive scenarios, such as areas with small load fluctuations; T h = 0.2 is used to capture significant changes and is suitable for power consumption scenarios with large fluctuations, such as industrial users or during peak load periods. By adjusting T within this range h , can flexibly adapt to the needs of different scenarios and ensure the accuracy and rationality of the recognition of change points. By setting a reasonable threshold range T h , can effectively distinguish significant change points in electricity consumption behavior from small fluctuations, and avoid false alarms or missed alarms. The dynamic adjustment of the threshold range is applicable to a variety of electricity consumption scenarios, including slight fluctuations in household users and drastic changes in industrial users, which enhances the versatility of the method. The recorded change point data provides support for subsequent analysis, such as drawing load change trend charts, identifying abnormal behavior patterns, etc. By accurately identifying change points and optimizing segmentation processing and feature extraction steps, the accuracy of short-term power load forecasting is improved. The calculation process is simple and efficient, and can quickly locate electricity consumption change points, supporting real-time data analysis and load management.

[0086] Based on the calculation result of ΔC, if ΔC>T h , then recording the change point includes comparing ΔC with a preset threshold value T h , if ΔC>T h ,Record time t as the change point, record the change type of the power consumption behavior pattern based on the change point, and update the statistical parameters of the power consumption behavior pattern after the change point.

[0087] According to the calculated change rate ΔC and the preset threshold T h When ΔC>T hWhen the current moment t is reached, it is determined that there is a significant change in the electricity consumption behavior pattern, and further records and analysis are performed. Based on the determination of the change point, information related to the change point is recorded, including the change point time t; the current power load P(t) and the negative P(t1) at the previous moment; the rate of change ΔC; and the current external environment (such as meteorological conditions and social activities). These records provide basic data for subsequent analysis of electricity consumption behavior patterns. According to the characteristics of the change point, combined with external conditions (such as load data trends and meteorological conditions), the type of change is identified. For example, a sudden increase in load: load data increases significantly in hot weather or during peak electricity consumption; a sudden drop in load: load data decreases significantly during holidays or during low electricity consumption periods; abnormal fluctuations: the rate of change of load data fluctuates abnormally compared with historical data.

[0088] Each type of change provides guidance for the optimization of power grid dispatch and prediction models. After identifying the change point, the statistical parameters related to the current power consumption behavior pattern are updated, such as the average load value: recalculate the average load level after the change point; change amplitude distribution: record the distribution of the change rate of the current behavior pattern; fluctuation frequency: count the frequency of occurrence of the change point under the behavior pattern. These statistical parameters can reflect the dynamic characteristics of power consumption behavior and provide support for the optimization of input features of deep learning models. By comparing the change rate with the threshold, significant change points of power consumption behavior can be efficiently identified to avoid misjudgment of slight fluctuations. The classified record of change points provides a clear basis for the analysis of power consumption behavior patterns, which helps to deeply understand the law of load changes. By updating the statistical parameters after the change point, the changing trend of power consumption behavior can be reflected in time, and the dynamic adaptability of the prediction model can be improved. The method can be applied to a variety of power consumption scenarios (such as residential power consumption, commercial power consumption or industrial power consumption), with high flexibility and promotion value.

[0089] Recording the change type of the electricity consumption behavior pattern based on the change point includes identifying the electricity consumption behavior pattern type before and after the change, classifying the change type of the electricity consumption behavior pattern based on the change point and the pattern matching result, and generating a pattern change type report.

[0090] Use cluster analysis or classification models to perform pattern recognition on the load data before and after the change point, extract the key features of each data segment (such as average load, fluctuation amplitude, load peak, etc.); match the recognition results with the defined pattern type to determine which type of power consumption behavior pattern before and after the change. Based on the conversion of the pattern at the change point, classify the change type of the power consumption behavior pattern. For example, when the load data shows a significant increase at the change point and matches the peak load characteristics, it is determined to be "switched from a stable load type to a peak load type"; combine historical records and external data (such as meteorological conditions or holiday data) to analyze the possible reasons for the pattern change and improve the classification accuracy. Comprehensive pattern change classification results generate a structured pattern change type report; the report content includes change point information: the time of the change point, the rate of change of power consumption; pattern change type: the pattern type before and after the change and the description of the change type; statistical analysis: statistical parameters such as the frequency distribution of the change point and the load change amplitude; external influencing factors: such as a sudden increase in load caused by high temperature weather, or a sudden decrease in load caused by holidays. The report can be used for subsequent optimization of power dispatch and load forecasting models. Through accurate identification of patterns before and after changes, the evolution characteristics of electricity consumption behavior can be clarified, providing a scientific basis for the classification of change points. Based on the change point information and combined with the pattern matching results, the change type of electricity consumption behavior pattern can be accurately classified, which is helpful for in-depth analysis of the causes of load fluctuations. The generated pattern change type report provides a comprehensive view of load changes and provides a basis for scheduling optimization and resource allocation of power systems. The method is applicable to a variety of scenarios, including load change pattern analysis of residential, commercial and industrial users, which enhances the practicality and adaptability of the method.

[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A short-term power load forecasting method based on deep learning, characterized in that: The method comprises: S1. Collect historical power load data and corresponding weather and holiday information; S2. Based on the historical power load data and the changes in the user's power consumption behavior pattern, the historical data is preprocessed to reduce data inconsistency. Specifically, the historical load data is smoothed, abnormal fluctuations and peak data are removed, and the power consumption data is seasonally decomposed to extract long-term trends, cyclical changes and random components. Data from different sources are aligned by timestamps and missing values ​​are filled. The formula for smoothing the historical load data is: S t =αY t +(1-α)S t-1 ; Among them, S t It represents the historical power load data value at a certain moment after smoothing, Y t represents the original power load data actually measured at a certain moment, α represents the smoothing factor, S t-1 It represents the result of smoothing the power load data at the previous moment, and t represents time; S3, building a deep learning model, the model input is processed historical power load data, weather and holiday information; S4. Use the training data set to train the constructed deep learning model and optimize the model parameters; S5. Apply the trained deep learning model to perform short-term power load forecasting and generate forecast results.

2. A short-term power load forecasting method based on deep learning according to claim 1, characterized in that: The S1 includes: The geographic information of the power load collection points is spatially matched with the meteorological data, and the sampling point data is spatially interpolated to estimate the meteorological conditions at other locations and generate a continuous spatiotemporal meteorological data table. Combined with the calendar data of holidays and weekdays, the load data and meteorological data are time-annotated to form a unified spatiotemporal sequence input. The calculation formula for spatial interpolation of the sampling point data is: Z(s)=μ(s)+∈(s); Among them, Z(s) represents a target value at the geographic space location s, s represents the geographic coordinate location of the power load collection point or meteorological station, μ(s) represents the overall trend or long-term stable value of data in a large range, and ∈(s) represents the random error term.

3. The short-term power load forecasting method based on deep learning according to claim 1 is characterized in that: The S3 includes: Construct a one-dimensional convolutional network, perform convolution operations on time series data, extract local features, and combine the bidirectional long short-term memory network to capture the bidirectional dependency of the time series. The one-dimensional convolution operation formula is: Among them, y[i] represents the eigenvalue or the i-th element in the feature sequence obtained by the one-dimensional convolution operation, x[i] represents the i-th data point in the original data sequence input to the deep learning model, w[k] represents the weight coefficient of the convolution kernel for the k-th position in the input sequence during the convolution operation, k represents the position index in the convolution kernel, K represents the size of the convolution kernel, and i represents the starting position index of the input sequence.

4. The method for short-term power load forecasting based on deep learning according to claim 1, characterized in that: The S4 includes: The historical load data is divided into multiple training tasks by time period, the changes in different power consumption scenarios are simulated, and the initial parameters of the deep learning model are adjusted. The specific formula is: Among them, θ * represents the final parameter value of the deep learning model after meta-learning optimization, θ represents the current parameter value of the deep learning model, and α represents the learning rate. represents the gradient operator, Represents the loss function for evaluating the overall performance of the model on different tasks.

5. The method for short-term power load forecasting based on deep learning according to claim 1, characterized in that: The S5 includes: Use the trained deep learning model to generate load forecast results, and use known observation data to calibrate the posterior probability of the forecast results. The specific formula is: Among them, P(X|E) represents the probability value of predicting short-term power load X under the condition of known observation data E, P(E|X) represents the probability of observation data E appearing under a certain power load X, P(X) represents the initial probability of power load X in the absence of observation data E, and P(E) represents the overall probability of observation data E appearing.

6. The method for short-term power load forecasting based on deep learning according to claim 1, characterized in that: The S2 further includes: Identify different electricity consumption behavior patterns of users through cluster analysis; Based on the time series analysis of the user's electricity consumption behavior pattern, the change points of the electricity consumption behavior pattern are detected; Identify and remove outliers from historical data; Based on the detected change points in electricity usage behavior patterns, historical data is processed in segments.

7. A short-term power load forecasting method based on deep learning according to claim 6, characterized in that: The time series analysis based on the user's electricity consumption behavior pattern detects the change point of the electricity consumption behavior pattern, including: Collect hourly power load data and corresponding meteorological information; Calculate the rate of change of the electricity consumption behavior pattern in each time period ΔC = (P(t) - P(t1)) / P(t1), where P(t) represents the electricity consumption at time t, and P(t1) represents the electricity consumption at time t1; Based on the rate of change ΔC, if ΔC>T h , it is considered that a change point in the electricity consumption behavior pattern has occurred in this time period, where T h is a pre-set threshold; Record the time of change points and electricity usage behavior patterns.

8. A short-term power load forecasting method based on deep learning according to claim 7, characterized in that: The calculation of the change rate of the electricity consumption behavior pattern in each time period ΔC=(P(t)-P(t1)) / P(t1) includes: Extract the power consumption data P(t) and P(t1) at time t and time t1; Calculate the power consumption change rate ΔC = (P(t) - P(t1)) / P(t1), where P(t) represents the power consumption at time t, and P(t1) represents the power consumption at time t1; Based on the calculation result of ΔC, if ΔC>T h , then record the change point; T h The value range of is [0.05, 0.2], which is used to distinguish significant change points from small fluctuations.

9. A short-term power load forecasting method based on deep learning according to claim 8, characterized in that: The calculation result based on ΔC, if ΔC>T h , then the record change points include: Compare ΔC with the preset threshold T h , if ΔC>T h ,Record time t as the change point, record the change type of the power consumption behavior pattern based on the change point, and update the statistical parameters of the power consumption behavior pattern after the change point.

10. A short-term power load forecasting method based on deep learning according to claim 9, characterized in that: The change types of the electricity consumption behavior pattern recorded based on the change point include: Identify the types of electricity consumption behavior patterns before and after the change, classify the change types of electricity consumption behavior patterns based on the change points and pattern matching results, and generate a pattern change type report.

Images