AI-based power load prediction method

Abstract

The application relates to the technical field of data reasoning, and discloses an AI-based power load prediction method, which comprises the following steps: compensating for missing values in target power data of a unified timestamp to obtain a standardized data set of a target power grid and adaptively extracting a multi-dimensional feature vector; predicting a historical power load based on a training set in a historical power data set; when the error value of the historical power load and a verification set in the historical power data set is lower than a load threshold, obtaining a load prediction weight; fitting a linear relationship between temperature change and the error value in the historical power data set, and taking the slope of the linear relationship as a compensation coefficient; predicting a power load condition of the target power grid, correcting load fluctuation of the power load condition based on the compensation coefficient; encoding the corrected power load condition into a power data packet, delivering the power data packet to the historical power data set in S2, and obtaining a target load prediction report; and the application can improve the accuracy of power load prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data reasoning, and particularly relates to an AI-based power load prediction method. BACKGROUND

[0002] In the operation and management of a power system, accurate power load prediction is a key link for ensuring stable power supply of a power grid and optimizing resource allocation. In the prior art, traditional power load prediction methods mostly rely on statistical models or single machine learning algorithms, and it is difficult to effectively handle complex nonlinear relationships and multi-dimensional influencing factors in power load data, such as dynamic variables such as weather changes, user power consumption behavior fluctuations, and economic policy adjustments, resulting in insufficient ability of the prediction model to capture actual load fluctuations, and a large deviation between the prediction results and the actual load data, making it difficult to meet the demand of fine management of the power grid.

[0003] In addition, with the rapid development of smart grids and new energy technologies, power load data presents high-dimensional and strong time-sensitive characteristics, and the existing technology has obvious defects in data feature extraction and fusion processing, and cannot fully mine the load change rules implied in massive real-time data. Especially in the face of sudden power consumption peaks or abnormal power consumption, the adaptability of the prediction model is poor, and the prediction results often lag behind or deviate greatly from the actual values, resulting in a lack of reliable basis for power dispatching decisions, and seriously affecting the economy and safety of the power grid operation. SUMMARY

[0004] The present application provides an AI-based power load prediction method, which mainly aims to solve the problem of low accuracy of power load prediction.

[0005] To achieve the above-mentioned purpose, the AI-based power load prediction method provided by the present application comprises:

[0006] S1. Compensate for missing values in the target power data of a unified timestamp to obtain a standardized data set of the target power grid, and adaptively extract a multi-dimensional feature vector of the standardized data;

[0007] S2. Predict the historical power load based on a training set in the historical power data set, and obtain the load prediction weight of the target power grid when the error value of the historical power load and a validation set in the historical power data set is lower than a load threshold;

[0008] S3. Fit the linear relationship between temperature changes in the historical power data set and the error value, and represent the slope of the linear relationship as a compensation coefficient of the historical power load;

[0009] S4. predicting the power load condition of the target power grid based on the load prediction weight and the multi-dimensional feature vector of the standardized data, and correcting the load fluctuation of the power load condition based on the compensation coefficient;

[0010] S5. encoding the corrected power load condition into a power data packet, and delivering the power data packet to the historical power data set in S2 to obtain a target load prediction report of the target power grid.

[0011] In a preferred embodiment, the compensation includes:

[0012] detecting an abnormal timestamp segment in the target power data;

[0013] inserting a missing value into the abnormal timestamp segment according to the space-time correlation of continuous missing;

[0014] replacing the isolated missing point in the abnormal timestamp segment with the mean value of adjacent data.

[0015] In a preferred embodiment, the multi-dimensional feature vector of the standardized data is extracted adaptively, including:

[0016] extracting a local fluctuation feature of the standardized data;

[0017] fusing the local fluctuation feature with a power information feature in the standardized data according to the dependency relationship of the local fluctuation feature to obtain a multi-dimensional feature vector of the standardized data.

[0018] In a preferred embodiment, the historical power load is predicted based on the training set in the historical power data set, and when the error value of the historical power load and the validation set in the historical power data set is lower than a load threshold, a load prediction weight of the target power grid is obtained, including:

[0019] dividing the historical power data set into a training set and a validation set;

[0020] initializing the weight of a time series prediction base model, generating an input sequence according to a time sliding window, inputting the input sequence into the time series prediction base model to obtain a prediction result of the training set;

[0021] updating the weight by the error value of the prediction result and the validation set to obtain an optimized weight of the time series prediction base model;

[0022] when the error value does not decrease, outputting the optimized weight when the error value does not decrease as the load prediction weight of the target power grid.

[0023] In a preferred embodiment, updating the weights using the error value between the prediction result and the validation set to obtain the optimized weights of the time series prediction base model includes:

[0024] Calculate the prediction error gradient of the error value, wherein the formula for calculating the prediction error gradient is as follows:

[0025] ;

[0026] In the formula, The error is... The weights of the time series prediction base model are... The fluctuation characteristics of the historical power load, The multidimensional characteristics of the historical power load, The input sequence, For partial derivative operators;

[0027] The historical learning rate of the time-series prediction base model is determined based on the historical electricity dataset, wherein the historical learning rate is calculated using the following formula:

[0028] ;

[0029] In the formula, For the first The learning rate updated next time. The initial learning rate is used for updating. The standard deviation of the historical power load is given. The number of updates to the time series prediction base model. The average of the historical power load is given. The historical power load;

[0030] The weights are updated based on the moving average of the gradient magnitude in the prediction error gradient and the historical learning rate.

[0031] In a preferred embodiment, fitting the linear relationship between temperature changes and error values ​​in the historical power data set includes:

[0032] Establish a mapping matrix between temperature changes and error values ​​in the historical power data set;

[0033] Based on the mapping matrix, a regression function is derived between the temperature change and the error value, wherein the regression function is as follows:

[0034] ;

[0035] In the formula, For the fluctuations in the historical power load, a thermal inertia factor of a power grid region, a temperature change, a time change of the temperature change, a base compensation constant;

[0036] the regression function as a linear relationship between the temperature change and the error value.

[0037] In a preferred embodiment, the slope of the linear relationship is characterized as a compensation coefficient of the historical power load, including:

[0038] extracting the fluctuation as a feature vector of the historical power load;

[0039] correcting the feature vector based on a thermal inertia parameter of a power device;

[0040] characterizing a module length of the corrected feature vector as a compensation coefficient.

[0041] In a preferred embodiment, the load fluctuation of the power load condition is corrected based on the compensation coefficient, including:

[0042] identifying a second derivative mutation point of a load curve in the power load condition;

[0043] segmenting a load interval with the mutation point as a boundary;

[0044] independently applying the compensation coefficient to the load interval to obtain a corrected power load condition.

[0045] In a preferred embodiment, the corrected power load condition is encoded as a power data packet and delivered to the historical power data set in S2, including:

[0046] encoding the corrected power load condition as a power load data set;

[0047] adding a time stamp check code to the power data set to generate a power data packet;

[0048] writing the historical power data set through a data bus cycle.

[0049] In a preferred embodiment, the target load prediction report of the target power grid is obtained, including:

[0050] segmenting the power data packet into data modules according to a preset time granularity;

[0051] adding a device topology check code to the data module to obtain a verification data block of the target power grid;

[0052] aggregate the verification data blocks to obtain a target load prediction report of the target power grid.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] 1. The present application can effectively capture the nonlinear relationship and dynamic change law in the load data by introducing artificial intelligence technology for multi-dimensional feature extraction and deep analysis of power load data, significantly improving the prediction accuracy of the model for load fluctuations in complex power consumption scenarios, making the prediction results closer to the actual power load curve, and providing more reliable decision basis for power grid dispatching.

[0055] 2. The present application can use the adaptive learning ability of AI algorithm to process and optimize the model in real time for high-dimensional and high-timeliness load data in smart grids, effectively solve the shortcomings of traditional methods in data feature fusion and sudden situation response, realize accurate prediction of power load, and improve the efficiency of power grid resource allocation, enhance the stability and economy of power grid operation, and meet the demand for high-precision load prediction in fine management of smart grids. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 The flowchart of the AI-based power load prediction method provided by an embodiment of the present application is shown.

[0057] The implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0058] It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0059] The AI-based power load prediction method provided by the embodiments of the present application. The execution subject of the AI-based power load prediction method includes but is not limited to at least one of the electronic devices that can be configured to execute the method provided by the embodiments of the present application, such as a server and a terminal. In other words, the AI-based power load prediction method can be executed by software or hardware installed in a terminal device or a server device. The server includes but is not limited to a single server, a server cluster, a cloud server, or a cloud server cluster, etc. The server can be a standalone server, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), and big data and artificial intelligence platforms, etc. Basic cloud computing services.

[0060] Reference Figure 1As shown, it is a flowchart of an AI-based power load prediction method provided by an embodiment of the present application. In this embodiment, the AI-based power load prediction method comprises:

[0061] S1. Compensate for missing values in target power data with unified timestamps to obtain a standardized dataset of the target power grid, and adaptively extract a multi-dimensional feature vector of the standardized data;

[0062] In an embodiment of the present application, compensating for missing values in target power data with unified timestamps comprises:

[0063] Detecting abnormal timestamp segments in the target power data;

[0064] Inserting missing values into the continuously missing abnormal timestamp segments according to spatial and temporal correlation;

[0065] Using the mean value of adjacent data to replace isolated missing points in the abnormal timestamp segments.

[0066] Adaptively extracting a multi-dimensional feature vector of the standardized data comprises:

[0067] Extracting local fluctuation features of the standardized data;

[0068] According to the dependency relationship of the local fluctuation features, the local fluctuation features are fused with power information features in the standardized data to obtain a multi-dimensional feature vector of the standardized data.

[0069] Specifically, the target power data is traversed, the timestamps are arranged in order, and the interval between adjacent timestamps is checked one by one.

[0070] Further, if the interval between a certain timestamp and the previous timestamp is significantly greater than the normal interval range, or there is a sudden interruption of the timestamp, this part of the timestamp segment is marked as an abnormal timestamp segment, thereby detecting abnormal timestamp segments in the target power data.

[0071] Further, for continuously missing abnormal timestamp segments, spatially close and temporally adjacent other power data records are found using spatial and temporal correlation. The trend of the change of these related data in the corresponding time period is observed, and reasonable missing values are inserted in the continuously missing abnormal timestamp segments according to this trend.

[0072] Further, for isolated missing points in the abnormal timestamp segments, adjacent power data before and after the isolated missing point is found.

[0073] Further, the values of the adjacent data before and after are added and then divided by 2 to obtain the mean value of the adjacent data, and the mean value is used to replace the value of the isolated missing point.

[0074] Specifically, the standardized data is divided into multiple continuous small data segments according to a certain length. For each small data segment, the fluctuation of the data value in the segment is observed, and the amplitude and frequency of the change of the data value from the peak to the trough and from the trough to the peak are recorded. The information describing the fluctuation of the data value in the small data segment is extracted as the local fluctuation feature of the standardized data.

[0075] Further, the mutual influence relationship between the extracted local fluctuation features is analyzed, such as whether the change of a certain local fluctuation feature will trigger the corresponding change of other features.

[0076] Further, the features in the standardized data that can reflect power information, such as voltage value, current size, etc. are found.

[0077] Further, according to the dependency relationship of the local fluctuation features, the local fluctuation features and the power information features are integrated together, each feature is regarded as a dimension, and a vector containing multiple dimension information is formed, so as to obtain the multi-dimensional feature vector of the standardized data.

[0078] In summary, by compensating for the missing values of the target power data with the unified timestamp and constructing the standardized data set, the interference of data defects on the prediction model can be effectively eliminated, the integrity and consistency of the input data are ensured, high-quality basic data support is provided for load prediction, prediction deviation caused by data missing is avoided, and the model training is more in line with the actual operation law of the power grid.

[0079] In summary, the adaptive extraction of the multi-dimensional feature vector of the standardized data can automatically mine the complex features such as the spatio-temporal correlation and user power consumption mode implied in the power load data based on AI algorithm, accurately capture the nonlinear law and dynamic trend of load change, enhance the representation ability of the model for load fluctuation under different power consumption scenarios, and thus significantly improve the accuracy and reliability of the power load prediction result, providing a more accurate basis for power grid dispatching decision.

[0080] S2. predicting historical power load based on a training set in a historical power data set, and obtaining a load prediction weight of the target power grid when an error value of the historical power load and a validation set in the historical power data set is lower than a load threshold;

[0081] In the embodiment of the present application, the load prediction weight of the target power grid is obtained when the error value of the historical power load and the validation set in the historical power data set is lower than the load threshold based on the training set in the historical power data set, which includes:

[0082] dividing the historical power data set into a training set and a validation set;

[0083] initializing weights of a time series prediction base model, generating an input sequence according to a time sliding window based on the training set, and inputting the input sequence into the time series prediction base model to obtain a prediction result of the training set;

[0084] updating the weights according to an error value of the prediction result and the validation set to obtain optimized weights of the time series prediction base model;

[0085] when the error value does not decrease, outputting the optimized weights when the error value does not decrease as the load prediction weights of the target power grid.

[0086] The updating the weights according to the error value of the prediction result and the validation set to obtain the optimized weights of the time series prediction base model comprises:

[0087] calculating a prediction error gradient of the error value, wherein a calculation formula of the prediction error gradient is as follows:

[0088] ;

[0089] In the formula, is the error, is the weight of the time series prediction base model, is a fluctuation feature of the historical power load, is a multi-dimensional feature of the historical power load, is the input sequence, is a partial derivative calculation symbol;

[0090] determining a historical learning rate of the time series prediction base model according to the historical power data set, wherein a calculation formula of the historical learning rate is as follows:

[0091] ;

[0092] In the formula, is an updated learning rate, is an initial updated learning rate, is a standard deviation of the historical power load, is a number of updates of the time series prediction base model, is a mean value of the historical power load, is the historical power load;

[0093] updating the weights according to a gradient amplitude moving average value in the prediction error gradient and the historical learning rate.

[0094] ​Specifically, the historical power dataset is divided into a training set and a validation set according to a certain proportion, and the data in the training set and the validation set are ensured to be representative in time sequence and data characteristics.

[0095] Further, the weights of the connections between the nodes in the time series prediction base model are assigned initial values.

[0096] Further, in the training set, a sliding window of a fixed time length is set, and the data in the sliding window is sequentially combined into input sequences starting from the beginning of the training set data.

[0097] Further, the input sequences are input into the time series prediction base model with initialized weights one by one, and the model outputs the prediction results corresponding to the training set after processing.

[0098] Further, the prediction results of the training set are compared with the actual data of the validation set, and the error value between the two is calculated.

[0099] Further, according to the error value, the initialized weights in the time series prediction base model are adjusted, and by changing the weight size, the model prediction result is closer to the actual data, so as to obtain the optimized weights.

[0100] Further, the process of inputting the training set into the model to obtain the prediction result, calculating the error value with the validation set, and updating the weights is continuously repeated, and the change of the error value calculated each time is observed.

[0101] Further, when the error value no longer decreases, that is, no matter how the weights are updated, the error value will not become smaller or remain stable, the optimized weights obtained at this time are output as the weights used for load prediction of the target power grid.

[0102] Specifically, after comparing the error value obtained by comparing the prediction results of the training set and the actual data of the validation set, the trend of the error value changing with the change of the weights in the time series prediction base model is analyzed.

[0103] Further, the change of the error value when the weights change in different directions and amplitudes is observed, so as to determine the change direction and degree of the error value with respect to the weights, and to calculate the prediction error gradient of the error value.

[0104] Further, from the historical power data set, the weight update situation in the previous model training process is analyzed. The amplitude and direction of the weight update in the previous training round are observed, and according to this information, a suitable value is determined as the historical learning rate of the time series prediction base model, which reflects the general degree of weight update in the historical training of the model.

[0105] Further, the gradient amplitudes in the statistical prediction error gradient are calculated, a certain number of recent gradient amplitudes are selected from the amplitudes, the amplitudes are added and averaged to obtain a moving average of the gradient amplitudes. In combination with the determined historical learning rate, the weight of the time series prediction base model is adjusted according to the gradient change trend and degree indicated by the moving average of the gradient amplitudes. If the average value shows that the gradient changes in a certain direction, the weight size is changed in the direction according to the amplitude corresponding to the historical learning rate.

[0106] Specifically, the calculation formula of the prediction error gradient is That is, the error is obtained by comparing the prediction result of the training set through the time series prediction base model with the actual data of the verification set, and the difference between the two is the error.

[0107] Further, The weight of the time series prediction base model is the initial setting value of the connection between each node in the model, which will be updated during the model training process.

[0108] Further, The historical power load fluctuation feature is extracted from the historical power data set by observing the fluctuation of the historical power load data value in a certain time period, recording the change amplitude and frequency of the data value from the peak to the trough and from the trough to the peak, and the like.

[0109] Further, The historical power load multi-dimensional feature is the integration of the local fluctuation feature of the historical power data and the power information feature according to their dependency relationship to form a content containing multiple dimensional information.

[0110] Further, The input sequence is obtained by setting a sliding window with a fixed time length in the training set divided from the historical power data set, and sequentially combining the data in the sliding window.

[0111] Further, the significance of this formula is to calculate the influence degree of the error on the weight of the time series prediction base model.

[0112] Further, by multiplying the influence of the error on the historical power load fluctuation feature, the influence of the historical power load fluctuation feature on the historical power load multi-dimensional feature, and the influence of the historical power load multi-dimensional feature on the input sequence in turn, the comprehensive influence of the error on the weight is obtained, that is, the prediction error gradient, which reflects how the change of the weight affects the error.

[0113] Further, when the prediction error gradient value is positive, it means that if the weight is increased, the error value will increase, and at this time, reducing the weight can reduce the error; when the prediction error gradient value is negative, it means that increasing the weight will reduce the error, and reducing the weight will increase the error.

[0114] Further, the greater the absolute value of the prediction error gradient, the greater the change in error caused by a small change in weight; the smaller the absolute value, the smaller the impact of weight change on error.

[0115] Specifically, The learning rate value used when the time series prediction base model is updated for the first time, in the process of model training and weight updating, each update will correspond to such a learning rate.

[0116] Further, The learning rate set when the time series prediction base model is initially updated, that is, the initial learning rate, which is artificially set before the model training starts.

[0117] Further, By calculating the square of the difference between each data in the historical power load data set and the average value of the historical power load, adding these square values and taking the average, and then taking the square root of this average, the dispersion degree of the historical power load data is obtained.

[0118] Further, Indicates the cumulative number of times the time series prediction base model is updated from the start of training to the current time, and each time the weight is updated, this value increases by 1.

[0119] Further, Add all historical power load data in the historical power data set, and then divide by the total number of data to obtain the average level of historical power load data.

[0120] Further, The historical power load data contained in the historical power data set.

[0121] Further, the meaning of this formula is to determine the appropriate learning rate for updating the weight of the time series prediction base model at each time according to the distribution characteristics of the historical power load data and the number of model updates.

[0122] Further, by comprehensively considering the mean and standard deviation of the historical power load, and combining the number of updates the model has made, the learning rate is dynamically adjusted to make the model more reasonably update the weight during training.

[0123] Further, as the number of updates of the time series prediction base model increases ​With the increase of the number of updates, the value of the denominator in the formula gradually increases.

[0124] Further, because the denominator is based on 1 plus a value related to the historical power load standard deviation and mean multiplied by the number of updates and then square root, so the denominator increases, the entire fraction value becomes smaller, resulting in the learning rate of the first update gradually decreases.

[0125] Further, as the model training progresses, the learning rate becomes smaller and smaller, which can allow the model to adjust the weights more finely in the later training period, avoiding missing the optimal solution due to the learning rate being too large.

[0126] In summary, by performing load prediction based on the training set of the historical power data set and taking the error value of the validation set lower than the load threshold as the judgment condition for weight optimization, the obtained target power grid load prediction weight can have good generalization ability and fitting accuracy, so that the model can accurately capture the change rule of the power load in the historical data scene, effectively reduce the deviation between the prediction result and the actual load, and lay a reliable weight foundation for subsequent real-time load prediction.

[0127] In summary, the mechanism realizes adaptive optimization of the load prediction weight through dynamic verification of the error threshold, can accurately model the spatio-temporal characteristics and nonlinear characteristics of the power load data, and can maintain stable prediction performance of the model in different power consumption scenarios, thereby significantly improving the reliability and accuracy of the power load prediction result, and providing more accurate decision basis for power grid operation and dispatching.

[0128] S3. fitting the linear relationship between temperature change and error value in the historical power data set, and representing the slope of the linear relationship as a compensation coefficient of the historical power load;

[0129] In the embodiments of the present application, fitting the linear relationship between temperature change and error value in the historical power data set comprises:

[0130] establishing a mapping matrix of temperature change and error value in the historical power data set;

[0131] solving a regression function of the temperature change and the error value based on the mapping matrix, wherein the regression function is as follows:

[0132] ;

[0133] wherein, is the fluctuation of the historical power load, is a grid area thermal inertia factor, is the temperature change, a time variation of the temperature change, a base compensation constant;

[0134] the regression function is taken as a linear relationship between the temperature change and the error value.

[0135] the slope of the linear relationship is taken as a compensation coefficient of the historical power load, comprising:

[0136] the fluctuation is extracted as a feature vector of the historical power load;

[0137] the feature vector is corrected based on a thermal inertia parameter of a power device;

[0138] a module length of the corrected feature vector is taken as the compensation coefficient.

[0139] Specifically, when a mapping matrix of temperature change and error value is established in the historical power data set, temperature change data and corresponding error values at each time point are collected from the historical power data set.

[0140] Further, the temperature change data are arranged in columns, and the error values are arranged in corresponding time sequence and also arranged in columns, forming a matrix containing two columns of data, one column being specific values of the temperature change and the other column being corresponding error values, so that a mapping matrix of the two is established.

[0141] Further, when a regression function of the temperature change and the error value is solved based on the mapping matrix, a linear regression method is adopted.

[0142] Further, the mean of the temperature change data and the mean of the error values in the mapping matrix are calculated, and then the deviation of each temperature change data from the mean and the deviation of each error value from the mean are calculated respectively. The corresponding temperature deviation and error deviation are multiplied, all the products are added to obtain a deviation product sum, and the square sum of the temperature deviation is calculated.

[0143] Further, the slope of the regression function is obtained by dividing the deviation product sum by the square sum of the deviation, and the intercept is obtained by subtracting the slope multiplied by the mean of the temperature change from the mean of the error value.

[0144] Further, the regression function is taken as a linear relationship between the temperature change and the error value, and the linear relationship indicates that there is a linear correlation between the temperature change and the error value.

[0145] Further, when the temperature change increases, if the slope of the regression function is positive, the error value will increase by an amplitude corresponding to the slope; if the slope is negative, the error value will decrease by an amplitude corresponding to the slope. Through this linear relationship, the corresponding error value can be predicted according to the value of the temperature change, or the influence degree of the temperature change on the error value can be analyzed.

[0146] Specifically, a certain length of continuous data segment is selected from historical power load data, the fluctuation of load values in the data segment is observed, the peak and valley values of the data are recorded, the difference of load values at adjacent time points, the change amplitude per unit time, and the frequency of data fluctuation are calculated, and these values that can describe the fluctuation characteristics of load data are arranged in order to form a vector, thereby extracting the fluctuation characteristics as the historical power load feature vector.

[0147] Further, the thermal inertia parameters of the power equipment are obtained, which reflect the influence of equipment temperature change on load and the delay characteristics of temperature change.

[0148] Further, according to the heat conduction and heat dissipation law embodied by the thermal inertia parameters, each component in the extracted feature vector is adjusted, such as the delay part of the load fluctuation caused by the thermal inertia of the equipment, the time position or amplitude size of the related fluctuation characteristics in the feature vector is corrected according to the time delay degree corresponding to the thermal inertia parameters, so as to correct the feature vector based on the thermal inertia parameters of the power equipment.

[0149] Further, the square value of each component in the corrected feature vector is calculated, all the square values are added to obtain a sum, and the square root of the sum is taken to obtain a value, which is the module length of the corrected feature vector.

[0150] Further, the module length is taken as a compensation coefficient, which represents the comprehensive strength of the load fluctuation characteristics considering the influence of the thermal inertia of the equipment, and can be used for subsequent compensation and adjustment of the power load prediction model or related calculation.

[0151] Specifically, As the fluctuation of historical power load, the continuous data segment is selected from the historical power data set, the peak and valley values of the data, the difference between adjacent time points, the change amplitude and the fluctuation frequency are recorded, and the values that describe the fluctuation characteristics of the load are arranged in order to form a feature vector.

[0152] Further, is the thermal inertia factor of the power grid area, which is obtained by analyzing the thermal inertia parameters of the power equipment in the area where the power grid is located, such as the heat conduction characteristics and the heat dissipation rate, combining the corresponding relationship between temperature change and load fluctuation in historical data, and statistically calculating or experimentally measuring, and is used to reflect the influence degree of regional temperature change on load fluctuation.

[0153] Further, is the temperature change amount obtained by subtracting the temperature values at different time points at the same position in the historical power data set, is the time length interval between the occurrence of the temperature change, both of which are directly obtained from the temperature monitoring records and time stamps of the historical data.

[0154] Further, The base compensation constant is a fixed value obtained by statistical analysis of historical power load fluctuations without considering temperature changes, which is used to compensate for the base error or initial fluctuations. It can be determined by calculating the mean or initial state of historical load data.

[0155] Further, the significance of this regression function is to describe the linear relationship between historical power load fluctuations and temperature change rate.

[0156] Further, the temperature change is divided by the time change when the change occurs to obtain the temperature change rate, multiplied by the grid area thermal inertia factor, and then added to the base compensation constant, which can obtain the corresponding historical power load fluctuation value, which is used to predict or explain the load fluctuation situation through the temperature change rate, and reflects the influence mechanism of temperature change speed on load fluctuation.

[0157] Further, when the temperature change rate is large, if the grid area thermal inertia factor is positive, the historical power load fluctuation will increase by proportion with the increase of temperature change rate; if is negative, will decrease with the increase of temperature change rate.

[0158] Further, the base compensation constant is a fixed value that does not change with the temperature change rate, and always plays a basic compensation role for load fluctuation. The trend of the entire regression function is determined by the positive and negative of the temperature change rate and .

[0159] In summary, by fitting the linear relationship between temperature change and error value in historical power data set, the influence of temperature factor on power load prediction error can be accurately quantified, and the key environmental variable of temperature can be converted into a calculable compensation parameter. The model can dynamically respond to the nonlinear effect of temperature fluctuation on load, effectively correct the prediction deviation caused by temperature change, and improve the matching accuracy of load prediction value and actual electricity consumption data.

[0160] In summary, by using the linear relationship slope as the compensation coefficient to integrate into the load prediction weight, the model can realize adaptive adjustment of temperature sensitivity, so that the weight parameter can be dynamically optimized according to the temperature change law, and the model's ability to capture load fluctuations under different climate conditions can be enhanced.

[0161] In summary, this mechanism can significantly reduce the prediction error and improve the stability and accuracy of load prediction, especially in extreme temperature or seasonal temperature difference scenarios, providing more realistic environmental decision basis for grid dispatching.

[0162] S4. predicting the power load condition of the target power grid based on the load prediction weight and the multi-dimensional feature vector, and correcting the load fluctuation of the power load condition based on the compensation coefficient;

[0163] In the embodiments of the present application, the step of correcting the load fluctuation of the power load condition based on the compensation coefficient comprises:

[0164] identifying the second derivative mutation points of the load curve in the power load condition;

[0165] segmenting the load interval based on the mutation points;

[0166] independently applying the compensation coefficient to the load interval to obtain the corrected power load condition.

[0167] Specifically, for the load curve in the power load condition, the second derivative of each point in the load curve is calculated first.

[0168] Further, the first derivative of the load values of two adjacent time points, i.e., the load change rate, is calculated first, and then the first derivative is differentiated to obtain the second derivative, which reflects the speed of change of the load change rate.

[0169] Further, the numerical change of the second derivative is observed, and a reasonable threshold is set. When the numerical value of the second derivative of a point suddenly exceeds or is lower than the threshold, the point is identified as a second derivative mutation point of the load curve. These mutation points indicate that the load change trend has changed sharply.

[0170] Further, the identified mutation points are used as the dividing points to divide the entire load curve into multiple different load intervals.

[0171] Further, each load interval is bounded by two adjacent mutation points. After division, the load curve in each load interval has a relatively consistent change trend and similar change characteristics, such as a monotonically increasing, decreasing, or relatively stable state between the two mutation points.

[0172] Further, for each segmented load interval, the compensation coefficient is applied individually. The compensation coefficient is obtained by calculating the norm of the corrected feature vector, and it represents the comprehensive strength of the load fluctuation characteristics considering the influence of device thermal inertia.

[0173] Further, in each load interval, each load data point in the interval is multiplied by the corresponding compensation coefficient, or the load data is adjusted according to the compensation coefficient, so that each load interval is compensated independently, and finally the corrected power load condition is obtained, so that the load data more accurately reflects the actual operating condition.

[0174] Overall, by fusing the load prediction weight with the multi-dimensional feature vector for power load prediction, the deep mining capability of AI algorithms for multi-dimensional features such as spatio-temporal correlation and user power consumption patterns in historical data can be fully utilized, enabling the model to accurately capture the nonlinear dynamic rules of power load data, and combining the pre-optimized load prediction weight to effectively improve the fitting accuracy of the prediction results to actual load fluctuations, providing more realistic power consumption scenario prediction data for power grid dispatching.

[0175] Overall, by introducing a compensation coefficient to correct the load fluctuations, the model can dynamically respond to the influence of environmental factors such as temperature on the prediction results, and by quantifying the linear relationship between temperature changes and error values, the model can adaptively adjust the prediction bias. In particular, in scenarios with dramatic temperature changes or seasonal transitions, the prediction error caused by environmental factors is significantly reduced, enhancing the stability and reliability of the load prediction results and further improving the decision-making accuracy of power grid operation and dispatching.

[0176] S5. The corrected power load is encoded into a power data packet and delivered to the historical power data set in S2 to obtain a target load prediction report for the target power grid.

[0177] In an embodiment of the present application, obtaining the target load prediction report for the target power grid comprises:

[0178] The power data packet is divided into data modules according to a preset time granularity.

[0179] An equipment topology check code is added to the data modules to obtain a verification data block for the target power grid.

[0180] The verification data block is aggregated to obtain the target load prediction report for the target power grid.

[0181] Specifically, the power data packet is divided into data modules according to a preset time granularity. First, the preset time granularity is determined, such as hours, days, or weeks, etc.

[0182] Further, according to this time granularity, the data in the power data packet is divided in chronological order, and the data in the same time period is combined together to form independent data modules. Each data module contains power data in the time period.

[0183] Further, an equipment topology check code is added to the data modules to obtain a verification data block for the target power grid. First, the equipment topology structure information of the target power grid is obtained, including the connection relationship and attributes of the equipment, etc.

[0184] Further, based on these topology information, a unique device topology verification code is generated. This verification code is added to each data module, combined with the data module, forming a verification data block containing data and verification code.

[0185] Further, the verification data blocks are aggregated to obtain the target load prediction report of the target power grid. All generated verification data blocks are collected and arranged in a certain logical order.

[0186] Further, the data in these verification data blocks are summarized and integrated to extract information related to load prediction.

[0187] Further, based on these information, analysis and prediction are performed to finally generate a complete target load prediction report of the target power grid, which contains the prediction results of the target power grid load.

[0188] In summary, the corrected power load conditions are encoded into standardized power data packets and delivered to the historical power data set. Through the data cycle iteration mechanism, the sample capacity and timeliness of the training data can be continuously expanded. The historical data set can dynamically absorb the latest load prediction correction results, forming a "prediction-correction-data update" closed-loop optimization process. This enhances the learning ability of the AI model for the load change law of the power grid, ensures that the model can optimize parameters based on more comprehensive and more realistic operating state data in subsequent training, and effectively improves the long-term accuracy and adaptability of load prediction.

[0189] In summary, the target load prediction report of the target power grid generated through the above data integration process can systematically present the corrected prediction results in a structured form. Not only can it intuitively reflect the real-time accuracy of load prediction and error correction effect, but also can provide decision-making basis for power grid dispatchers through comparison and analysis of historical data and correction data, including load trend, error compensation strategy and other multi-dimensional information. This makes the prediction report more suitable for the needs of fine management of the power grid, and further improves the stability and resource allocation efficiency of the power system operation.

[0190] In the several embodiments provided in the present application, it should be understood that the disclosed method can be implemented in other ways.

[0191] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0192] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. The artificial intelligence is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An AI-based power load forecasting method, characterized by, The method comprises: S1. Compensate missing values in target power data with unified timestamp to obtain a standardized data set of the target power grid, and adaptively extract a multi-dimensional feature vector of the standardized data; S2. Predict historical power load based on a training set in a historical power data set, and obtain a load prediction weight of the target power grid when an error value of the historical power load and a validation set in the historical power data set is lower than a load threshold; S3. Fit a linear relationship between temperature change and the error value in the historical power data set, and represent a slope of the linear relationship as a compensation coefficient of the historical power load, comprising: extracting fluctuations of the historical power load as a feature vector of the historical power load; correcting the feature vector based on thermal inertia parameters of power equipment; representing a module length of the corrected feature vector as the compensation coefficient; S4. Predict a power load condition of the target power grid based on the load prediction weight and the multi-dimensional feature vector, and correct load fluctuations of the power load condition based on the compensation coefficient, comprising: identifying a second derivative mutation point of a load curve in the power load condition; segmenting a load interval based on the mutation point; independently applying the compensation coefficient to the load interval to obtain a corrected power load condition; S5. Encode the corrected power load condition into a power data packet and deliver it to the historical power data set in S2 to obtain a target load prediction report of the target power grid. 2.The AI-based electric power load prediction method of claim 1, wherein, The compensation of missing values in target power data with unified timestamp comprises: detecting abnormal timestamp fragments in the target power data; inserting missing values into the abnormal timestamp fragments with continuous missing according to space-time correlation; using mean values of adjacent data to replace isolated missing points in the abnormal timestamp fragments. 3.The AI-based electric power load prediction method of claim 1, wherein The adaptive extraction of the multi-dimensional feature vector of the standardized data comprises: extracting local fluctuation features of the standardized data; fusing the local fluctuation features with power information features in the standardized data according to a dependency relationship of the local fluctuation features to obtain a multi-dimensional feature vector of the standardized data. 4.The AI-based electric power load prediction method of claim 1, wherein The prediction of the historical power load based on the training set in the historical power data set, and the obtaining of the load prediction weight of the target power grid when an error value of the historical power load and a validation set in the historical power data set is lower than a load threshold, comprises: dividing the historical power data set into a training set and a validation set; initializing a weight of a time series prediction base model, generating an input sequence from the training set according to a time sliding window, inputting the input sequence into the time series prediction base model to obtain a prediction result of the training set; updating the weight through an error value of the prediction result and the validation set to obtain an optimized weight of the time series prediction base model; when the error value does not decrease, outputting the optimized weight when the error value does not decrease as the load prediction weight of the target power grid. 5.The AI-based electric power load prediction method of claim 4, wherein The updating of the weight through the error value of the prediction result and the validation set to obtain the optimized weight of the time series prediction base model comprises: calculating a prediction error gradient of the error value, wherein a calculation formula of the prediction error gradient is as follows: ； wherein, is the error, is a weight of the timing prediction base model, is a fluctuation feature of the historical power load, is a multi-dimensional feature of the historical power load, is the input sequence, is a partial derivative calculator; According to the historical power data set, a historical learning rate of the time series prediction base model is determined, wherein a calculation formula of the historical learning rate is as follows: ； In the formula, is the learning rate of the first update, is the learning rate of the initial update, is the standard deviation of the historical power load, is the number of updates of the time series prediction base model, is the mean of the historical power load, is the historical power load; According to the gradient amplitude moving average value in the prediction error gradient and the historical learning rate, the weight is updated. 6.The AI-based electric power load prediction method of claim 1, wherein, The fitting of the linear relationship between the temperature change in the historical power data set and the error value comprises: A mapping matrix of the temperature change and the error value in the historical power data set is established; Based on the mapping matrix, a regression function of the temperature change and the error value is solved, wherein the regression function is as follows: ； wherein is a fluctuation of the historical power load, is a grid zone thermal inertia factor, is the temperature change, is a time change at which the temperature change occurs, is a base compensation constant; The regression function is taken as the linear relationship between the temperature change and the error value. 7.The AI-based electric power load prediction method of claim 1, wherein The modified power load condition is encoded into a power data packet, and is delivered to the historical power data set in S2, which comprises: The modified power load condition is encoded into a power load data set; A time stamp check code is added to the power data set to generate a power data packet; The historical power data set is written through a data bus cycle. 8.The AI-based electric power load prediction method of claim 1, wherein, The target load prediction report of the target power grid is obtained, which comprises: The power data packet is divided into data modules according to a preset time granularity; A device topology check code is added to the data modules to obtain a verification data block of the target power grid; The verification data block is aggregated to obtain the target load prediction report of the target power grid.

Images