Power demand prediction method and system based on big data analysis

Abstract

This invention discloses a method and system for electricity demand forecasting based on big data analysis, belonging to the field of power big data analysis and load forecasting technology. It addresses the problems of multi-step rolling error accumulation and large peak-valley load forecasting deviations in traditional electricity demand forecasting. This invention collects and standardizes multi-source data on historical load, meteorological, date characteristics, and electricity consumption structure, jointly identifying and correcting abnormal data. Through error propagation characteristic analysis, a comprehensive error propagation index is constructed. Based on wavelet decomposition, hierarchical forecasting of trend, periodic, and fluctuation components is performed, and a dynamic weight allocation strategy is designed in conjunction with the error propagation index. An error compensation and confidence decay mechanism is introduced, and rolling updates and segmented verification corrections are performed, significantly improving the stability and accuracy of electricity load forecasting.

Description

Technical Field

[0001] This invention relates to the field of power big data analysis and load forecasting technology, specifically to a power demand forecasting method and system based on big data analysis. Background Technology

[0002] Electricity demand forecasting is a crucial foundation for grid dispatching, generation planning, and electricity market transactions. Accurate medium- and long-term load forecasting is of great significance for ensuring the safe and economical operation of the power grid. With the advancement of smart grid construction, the power system has accumulated massive amounts of historical load data, meteorological data, and electricity consumption behavior data, providing data support for big data prediction methods. Existing forecasting methods mainly rely on time series analysis, regression analysis, neural networks, and deep learning. Deep learning methods for big data have outstanding nonlinear fitting capabilities and perform well in short-term forecasting, but their performance degrades significantly in the medium and long term. Most existing methods use autoregressive rolling forecasting, where the forecasting error is propagated and amplified step by step. The single-step forecasting error is about 3% to 5%, and the error can reach 15% to 25% by the 7th day. The error exceeds 30% during peak and valley load periods, rendering the forecasting results useless. The core reasons for error accumulation are: on the one hand, the long-term prediction input changes from the true value to the predicted value, and the data distribution deviates from the training sample; on the other hand, the error propagation characteristics of different components are not distinguished and processed, and the error of high-frequency fluctuating components will quickly pollute the overall prediction results; although some studies have tried to avoid gradual rolling through direct multi-output models, such methods ignore the temporal dependencies between future moments, resulting in insufficient continuity and physical rationality of the prediction sequence.

[0003] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of multi-step rolling error accumulation and large peak-valley load forecasting deviation in traditional power demand forecasting, and to propose a power demand forecasting method and system based on big data analysis.

[0005] The objective of this invention can be achieved through the following technical solutions: Electricity demand forecasting methods based on big data analytics include: S1. Historical data acquisition and preprocessing: The historical load, weather, date characteristics and electricity consumption structure data of power equipment are acquired through the data acquisition unit, and missing value filling, abnormal data correction and standardization are performed to form a standardized load time series dataset. S2. Error propagation characteristic analysis: Obtain the prediction error and error amplification coefficient, extract the load fluctuation index, construct the error propagation sensitivity network, and combine the temperature-sensitive load ratio and peak-valley fluctuation amplification factor to obtain the comprehensive error propagation index through weighted fusion. S3. Generation of hierarchical prediction strategies: Based on the analysis results of error propagation characteristics, the prediction task is decomposed into trend component prediction, periodic component prediction and fluctuation component prediction, and independent prediction strategies for each component are designed. S4. Error Compensation Prediction Execution: The weighted fusion component prediction values ​​are used to obtain preliminary prediction values. An error estimation fitting network is constructed, and a confidence decay mechanism is introduced for error compensation. The error propagation characteristic parameters are iterated based on the triggered rolling update mechanism. S5. Prediction Result Verification and Correction: Perform segmented verification of the prediction results, save parameters, perform local optimization or configuration rollback based on the error results, and perform rolling correction based on the actual data.

[0006] Furthermore, the specific operation steps of S2 are as follows: Extract the actual load sequence of consecutive days from historical load data, construct the load time series matrix, select sample intervals from historical load data, and simulate multi-step rolling forecasts; Set the prediction step size, calculate the prediction error and error amplification factor for each step, and determine the error accumulation phenomenon; The load standard deviation sequence is extracted to obtain the load fluctuation index. The error amplification factor is correlated with the load fluctuation index to construct an error propagation sensitivity network. Temperature, humidity, and wind speed parameters are extracted from meteorological data to calculate the daily temperature variation and the proportion of temperature-sensitive loads. Identify peak load times, statistically analyze peak and valley patterns on weekdays and non-weekdays, and calculate peak and valley fluctuation amplification factors; fit the error propagation coefficients for different date types to obtain error propagation correction factors; After normalizing the error propagation sensitivity, the proportion of temperature-sensitive loads, the peak-valley fluctuation amplification factor, and the error propagation correction factor, the comprehensive error propagation index is obtained through a weighted fusion formula.

[0007] Furthermore, the specific operation steps of S3 are as follows: The preprocessed historical load sequence was decomposed into three levels using the Daubechies wavelet basis function to extract low-frequency trend components, mid-frequency periodic components, and high-frequency fluctuation components. The trend component is smoothed by moving average with a dynamic window length, and the predicted value of the trend component is obtained by least squares linear regression fitting. Frequency domain analysis of the periodic components is performed using fast Fourier transform to identify the dominant period and construct a period superposition model to obtain the predicted values ​​of the periodic components. Long Short-Term Memory (LSTM) networks are used to fit and predict short-term fluctuations in the fluctuation components. Once the predicted value of the fluctuation component exceeds the set short-term step size, it is set to zero. Based on the comprehensive error propagation index, a dynamic weight allocation strategy is designed to adaptively adjust with the prediction step size, and dynamic weight allocation is performed on each component to form a hierarchical independent prediction strategy.

[0008] Furthermore, the specific operation steps of S4 are as follows: The predicted values ​​of trend component, cycle component and fluctuation component obtained from the hierarchical prediction are weighted and fused based on the corresponding dynamic weights to obtain the preliminary prediction value; An error estimation fitting network is constructed, with the actual error of the previous prediction step size as the input feature. A sliding window collects historical errors from multiple steps to form fitting samples. The time-series fitting is extrapolated through an autoregressive model to obtain the current step size error estimate. A confidence decay mechanism is introduced to calculate the confidence level corresponding to the prediction step size; based on the prediction confidence level, the preliminary prediction value and the benchmark prediction value are weighted and fused to obtain the compensated prediction value; During the prediction process, the time points of the acquired actual load data are monitored in real time, and the newly added actual load data are statistically analyzed point by point. When the cumulative newly connected actual load data reaches the preset time, a rolling update is triggered, and the error propagation-related parameters are recalculated and the weight coefficients and confidence decay coefficients of the autoregressive model are updated.

[0009] Furthermore, the specific operation steps of S5 are as follows: The compensated predicted value sequence is divided into three validation segments—short-term, medium-term, and long-term—based on the prediction step size, and an accuracy threshold is set for each validation segment. Collect actual load data and calculate the prediction error index segment by segment; determine the prediction result based on the prediction error index of each verification segment and the corresponding set accuracy threshold. If the prediction result is deemed to have passed the verification, the multi-source parameters, including wavelet decomposition parameters, linear extrapolation equation coefficients, and error estimation fitting network parameters, are saved simultaneously to form a prediction configuration scheme. If the prediction result is determined to be locally optimized, then locate the current validation segment and adjust the error propagation correction factor and error estimation weight; If the prediction is determined to be unsuccessful, the configuration rollback mechanism is triggered to automatically restore the prediction parameters and configuration scheme that passed the previous round of verification, and an abnormal prediction log is generated. After verification, based on the deviation between the actual load value and the compensated predicted value, the exponential smoothing method is used to perform rolling correction on the subsequent prediction sequence. The correction weight decreases to zero as the prediction step size increases, and the load time series dataset is updated with the corrected prediction results for subsequent multi-step rolling prediction.

[0010] A second aspect of the present invention provides a power demand forecasting system based on big data analysis, comprising: Data acquisition and preprocessing module: By acquiring historical power load, meteorological, date characteristics and power consumption structure data, and performing missing value imputation, abnormal data correction and standardization processing, a standardized load time series dataset is generated; Error propagation characteristics analysis module: Constructs a load time series matrix, calculates prediction error, error amplification factor and load fluctuation index, constructs an error propagation sensitivity network, and combines the proportion of temperature-sensitive loads and peak-valley fluctuation amplification factor to obtain a comprehensive error propagation index; Hierarchical prediction strategy generation module: Performs wavelet decomposition on the load sequence to predict the trend, periodic and fluctuation components respectively, and assigns dynamic weights based on the comprehensive error propagation index to form a hierarchical independent prediction strategy; Error Compensation Prediction Execution Module: Obtains preliminary prediction values ​​by weighted fusion of the prediction values ​​of each component, estimates the error by autoregressive model, and performs error compensation by combining confidence decay. Prediction result verification and correction module: Verify prediction accuracy by dividing the prediction into short, medium and long term, perform parameter saving, local optimization or configuration rollback based on prediction results, and use exponential smoothing rolling correction to update the load time series dataset.

[0011] Compared with the prior art, the beneficial effects of the present invention are: This invention identifies and corrects anomalous data by standardizing and preprocessing multi-source data on historical load, meteorological, date characteristics, and power consumption structure; performs error propagation characteristic analysis to construct a comprehensive error propagation index to quantify the error accumulation law; performs hierarchical independent prediction of load trend, periodic, and fluctuation components based on wavelet decomposition, and designs a dynamic weight allocation strategy in combination with error propagation characteristics; introduces error compensation and confidence decay mechanisms to suppress long-step prediction bias, and continuously optimizes through segmented verification and rolling correction, significantly improving the accuracy and stability of multi-step rolling prediction, and enhancing the continuity and practicality of prediction results. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example: like Figure 1As shown, the power demand forecasting method based on big data analysis includes historical data acquisition and preprocessing, error propagation characteristic analysis, hierarchical forecasting strategy generation, error compensation forecasting execution, and forecasting result verification and correction.

[0015] S1. Historical Data Acquisition and Preprocessing: Historical load data, meteorological data, date characteristic data, and electricity consumption structure data of power equipment are acquired through the data acquisition unit, including: Historical load data includes point-by-point load power sequences of buses, typical feeders and regional summaries at each voltage level. The time resolution is uniformly 15 minutes, and the collection duration covers no less than two consecutive calendar years, forming a 96-point daily load time series. Meteorological data are collected synchronously for hourly temperature, humidity, wind speed, precipitation, and light intensity in the corresponding area, and spatiotemporally aligned with the load time series. Date feature data includes Gregorian calendar dates, weekday identifiers, statutory holidays, adjusted workdays, and extreme weather warning markers, and date type feature vectors are constructed; Electricity consumption structure data is broken down into sub-items of electricity consumption types such as residential, commercial, industrial, agricultural and charging piles, forming a multi-dimensional electricity consumption structure matrix; Preprocessing of multi-source heterogeneous data: Linear interpolation was used to fill in load and meteorological data with no more than three consecutive missing time points. For data with more than three consecutive missing time points, the historical average of the same period under similar weather conditions and date type was used as a substitute. Anomalies were identified by combining the 3σ criterion and the isolated forest algorithm, including sudden increases and decreases in load, data overflow, negative values, and data collection breakpoints. For load points identified as abnormal, a sliding window weighted average was used for correction. The length of the sliding window was the number of time points corresponding to 24 hours, and the weight decreased with time distance. The preprocessed data is standardized to map the load values ​​to the [0, 1] interval. At the same time, timestamp unification, time zone correction and feature alignment are completed to form a standardized load time series dataset.

[0016] S2. Error propagation characteristics analysis: Extracting continuous data from historical load data The actual load sequence of the day, where The actual load sequence is divided into 96 points based on time granularity to construct a load time series matrix; The time period of the actual load value collected in the preprocessed historical load data is selected as the sample interval; a multi-step rolling forecasting process is simulated based on the sample interval. Set the prediction step size range to 1 to Step, among which The number of time points corresponding to the target prediction number of days; using the formula Calculations are performed to obtain the first step-by-step simulation prediction error ,in, This represents the total number of prediction windows used in statistical calculations. Indicates the first In the prediction window, corresponding to the _th The predicted load value of the step, Indicates the first In the prediction window, corresponding to the _th The actual load value of the step; Based on the simulation prediction errors at each step, using the formula The error amplification factor is calculated, where, This represents the baseline value for single-step prediction error; When the error amplification factor is greater than the preset threshold, it indicates that there is an error accumulation phenomenon. Obtain the standard deviation sequence of historical load data, based on time. The load fluctuation index is obtained by comparing the load standard deviation of the same period in history with the load mean of the same period in history. The correlation analysis between the error amplification factor and the load fluctuation index was performed using the formula. Construct an error propagation-sensitive network, in which, Indicates the first Sensitivity to error propagation in step prediction The fluctuation impact coefficient is determined based on fitting historical data. Indicates the first The average load fluctuation index corresponding to the predicted time step; Simultaneously, temperature, humidity, and wind speed parameters are extracted from meteorological data, and the daily temperature variation is calculated using sliding window technology and a formula. The proportion of temperature-sensitive loads was calculated. ,in, Representing temperature sequence With load sequence covariance, Represents the covariance function. The standard deviation of a temperature series This represents the standard deviation of the load series; Identify peak load moments in historical data and analyze the patterns of peak and trough occurrences on weekdays and non-weekdays: Based on the 96-point daily load time series, the load values ​​at each time point are traversed day by day. The time of the maximum daily load is marked as the peak load time, and the time of the minimum load is marked as the valley load time. The distribution of peak load time and valley load time for weekdays, weekends and holidays within a continuous historical period are statistically analyzed to form the probability density distribution of peak and valley time periods. The peak and valley times under the same date type are clustered to obtain typical peak and valley time period intervals, and the load mean, load fluctuation rate and occurrence frequency corresponding to each time period are recorded. Through formula The peak-valley fluctuation amplification factor is calculated. ,in, This indicates the number of all peak load times selected from historical data. Indicates the first The prediction error corresponding to each peak time. Indicates the first The prediction error of a peak load period corresponding to a sample of peak times; Based on date feature data, weekday, weekend, and holiday types are identified, a mapping table between date types and load patterns is constructed, and the error propagation coefficients under different date types are fitted by the least squares method to obtain the error propagation correction factor. After normalizing the error propagation sensitivity, the proportion of temperature-sensitive loads, the peak-valley fluctuation amplification factor, and the error propagation correction factor, the comprehensive error propagation index is obtained through a weighted fusion formula.

[0017] S3, Generation of hierarchical prediction strategies: The preprocessed historical load sequence was decomposed using wavelet decomposition with Daubechies wavelet basis function. The number of decomposition layers was set to 3, and low-frequency components were extracted as trend components, mid-frequency components as periodic components, and high-frequency components as fluctuation components. For the trend component, a moving average smoothing process is used, and the window length is dynamically adjusted based on the prediction step size, using the formula... The window length for smoothing the trend components is calculated. , Describes the minimum value function. Represents the maximum value function. This represents the window adjustment factor, determined based on a fit of historical load data. This represents the upper limit threshold for the window length. The lower limit threshold representing the window length; Based on the smoothed trend sequence, the least squares method is used to perform linear regression fitting, and a linear extrapolation equation is constructed to obtain the predicted value of the trend component. The linear extrapolation equation is: ,in, Indicates the trend component in the th... The predicted value corresponding to each prediction step size. Indicates the slope of the trend, representing the rate of change of the trend components. The intercept represents the baseline level of the trend component; For the periodic components, the historical load sequence is analyzed in the frequency domain by fast Fourier transform to identify the dominant periodic components. Based on the average amplitude of all frequency components, frequency components with amplitudes greater than a preset threshold are selected to form the dominant frequency set. Based on the identified dominant cycle, a cycle superposition model is constructed, using the formula... Calculations are performed to obtain the predicted time. Predicted value of periodic components ,in, Represents the dominant frequency set, Represents frequency The corresponding amplitude, Represents frequency The corresponding phase angle, Pi is a constant. Represents the cosine function; For the fluctuation component, a long short-term memory network is used to fit and predict short-term fluctuations, but only effective predictions are made for the next 1-3 steps. After 3 steps, the predicted value of the fluctuation component is set to 0 to avoid error accumulation caused by high-frequency fluctuations in long-step prediction. Based on the comprehensive error propagation index, a dynamic weight allocation strategy that adaptively adjusts with the prediction step size is designed. The weights of the trend component, periodic component, and fluctuation component are calculated using the following formulas: Through formula The trend component weight values ​​are calculated. ; Through formula The periodic component weight values ​​are calculated. ; Through formula The weight values ​​of the fluctuation components are calculated. ; in, Indicates the maximum prediction step size; The dynamic weight allocation strategy ensures that as the prediction step size increases, the weight of the trend component increases monotonically, the weight of the periodic component decreases monotonically, and the weight of the fluctuation component decreases monotonically. Moreover, the weight change pattern matches the comprehensive error propagation index, forming a hierarchical independent prediction strategy that adapts to the error propagation characteristics of multi-step prediction.

[0018] S4. Error Compensation Prediction Execution: The predicted values ​​of trend component, periodic component, and volatility component generated by the hierarchical prediction strategy are weighted and fused with dynamic weight values ​​at the corresponding step size to obtain the preliminary prediction value. ; Construct an error estimation fitting network, using the actual error generated by the previous prediction step size as the input feature, and collect the most recent error through a sliding window. The historical errors of each step constitute the fitted samples, where, The values ​​are taken from 5 to 10 to form an error sequence. Based on the autoregressive model, the error sequence is time-series fitted and extrapolated to output the first... Error estimate corresponding to step ; A confidence decay mechanism is introduced to adapt to the objective law that the reliability of multi-step prediction gradually decreases as the prediction step size increases. This is achieved through the formula... Calculations are performed to obtain the first The prediction confidence level corresponding to each step of the prediction, where... Indicates the initial confidence level. Represents the natural constant. Indicates the confidence decay coefficient; The initial predicted value is fused with the baseline value after error compensation based on the prediction confidence level to obtain the compensated predicted value. The calculation formula is as follows: ,in, This represents the baseline forecast value, which is the median of the load at the same time in the same historical period. During the forecasting process, the time points of the acquired actual load data are monitored in real time, and the newly added actual load data are statistically analyzed point by point. When the cumulative newly connected actual load data reaches the preset time, the rolling update mechanism is triggered. The newly added actual load data is added to the historical dataset, and the error propagation characteristics related parameters such as error amplification coefficient, load fluctuation index, error propagation sensitivity, and comprehensive error propagation index are recalculated iteratively. At the same time, the weight coefficient and confidence decay coefficient of the autoregressive model are updated to achieve dynamic adaptive optimization of hierarchical forecasting and error compensation, and continuously suppress the error accumulation effect in the multi-step rolling forecasting process.

[0019] S5. Validation and correction of prediction results: By analyzing steps 1 to 2 After step compensation, the predicted value sequence is segmented for verification. The compensated predicted value sequence is divided into three verification segments: short-term, medium-term and long-term, according to the prediction step size. Precision thresholds are set for different verification segments, namely the first threshold for the short-term verification segment, the second threshold for the medium-term verification segment, and the third threshold for the long-term verification segment. Collect actual load data, calculate the prediction error index segment by segment, and use the formula The prediction error index is calculated. ,in, This represents the verification segment identifier, with values ​​from 1 to 3, corresponding to short-term, medium-term, and long-term respectively. Indicates the first The total number of time points contained in each verification segment Indicates the first Forecast load values ​​at each time point Indicates the first The actual load value corresponding to each time point; If the prediction error index of all verification segments is less than the corresponding set accuracy threshold, the prediction result is deemed to have passed the verification. The current wavelet decomposition parameters, linear extrapolation equation coefficients, periodic superposition model parameters, dynamic weight allocation coefficients, and error estimation fitting network parameters are saved simultaneously to form a stable and usable prediction configuration scheme. If the prediction error index of a single verification segment exceeds the corresponding accuracy threshold, and the excess is within the preset controllable range, it is determined that the prediction result is locally optimized. The time interval with large error in the current verification segment is located, and the load fluctuation index, temperature-sensitive load ratio and date type features of the corresponding time period are extracted. The error propagation correction factor and error estimation weight in the current scenario are adjusted to improve the error compensation accuracy. If the prediction error index of two or more verification segments exceeds the corresponding accuracy threshold, or the error of any verification segment exceeds the preset controllable range, the prediction is deemed to have failed, triggering the configuration rollback mechanism. The prediction parameters and configuration scheme that passed the previous round of verification are automatically restored, and the meteorological data, date characteristic data, power consumption structure data and comprehensive error propagation index at the time of prediction failure are recorded simultaneously to form an abnormal prediction log, providing data support for subsequent model iterations. After verifying the prediction results, rolling correction is performed based on the newly collected actual load data. The deviation between the actual load value and the compensated prediction value is used as the correction basis. The exponential smoothing method is used to correct the subsequent prediction sequence. The correction weights are allocated sequentially from small to large based on the prediction step size, so that the correction weights monotonically decrease and gradually decay to zero as the prediction step size increases. The prediction results after rolling correction replace the original compensated prediction values, are synchronously updated to the load time series dataset, and used for subsequent multi-step rolling predictions.

[0020] A power demand forecasting system based on big data analytics includes: Data acquisition and preprocessing module: By collecting historical load data of power equipment, meteorological data, date characteristic data and power consumption structure data, it performs missing value filling, abnormal data identification and correction, and standardization mapping processing on multi-source heterogeneous data to generate a standardized load time series dataset; Error propagation characteristic analysis module: Based on historical load sequences, a load time series matrix is ​​constructed, multi-step rolling forecasts are simulated and forecast errors and error amplification factors are calculated, load fluctuation index is extracted, an error propagation sensitivity network is constructed, the proportion of temperature-sensitive loads and peak-valley fluctuation amplification factors are calculated, and a comprehensive error propagation index is obtained by fusion. Hierarchical prediction strategy generation module: Performs three-level wavelet decomposition on the load sequence to extract trend, period, and fluctuation components. Component prediction is performed using linear regression, period superposition, and long short-term memory network, respectively. Dynamic weights are assigned based on the comprehensive error propagation index to form a hierarchical independent prediction strategy. Error Compensation Prediction Execution Module: Obtains preliminary prediction values ​​by weighted fusion of prediction values ​​of each component, fits historical errors through an autoregressive model and outputs error estimates, introduces a confidence decay mechanism for error compensation, and triggers rolling updates based on newly added actual load data; Prediction result verification and correction module: Divide the compensated prediction value sequence into three segments, short, medium and long term, for accuracy verification. Based on the prediction results, perform parameter saving, local optimization or configuration rollback, perform rolling correction using exponential smoothing, update the load time series dataset and use it for subsequent multi-step rolling prediction.

[0021] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A power demand forecasting method based on big data analysis, characterized in that, include: S1. Historical data acquisition and preprocessing: The historical load, weather, date characteristics and electricity consumption structure data of power equipment are acquired through the data acquisition unit, and missing value filling, abnormal data correction and standardization are performed to form a standardized load time series dataset. S2. Error propagation characteristic analysis: Obtain the prediction error and error amplification coefficient, extract the load fluctuation index, construct the error propagation sensitivity network, and combine the temperature-sensitive load ratio and peak-valley fluctuation amplification factor to obtain the comprehensive error propagation index through weighted fusion. S3. Generation of hierarchical prediction strategies: Based on the analysis results of error propagation characteristics, the prediction task is decomposed into trend component prediction, periodic component prediction and fluctuation component prediction, and independent prediction strategies for each component are designed. S4. Error Compensation Prediction Execution: The weighted fusion component prediction values ​​are used to obtain preliminary prediction values. An error estimation fitting network is constructed, and a confidence decay mechanism is introduced for error compensation. The error propagation characteristic parameters are iterated based on the triggered rolling update mechanism. S5. Prediction Result Verification and Correction: Perform segmented verification of the prediction results, save parameters, perform local optimization or configuration rollback based on the error results, and perform rolling correction based on the actual data.

2. The power demand forecasting method based on big data analysis according to claim 1, characterized in that, The specific operation steps of S2 are as follows: Extract the actual load sequence of consecutive days from historical load data, construct the load time series matrix, select sample intervals from historical load data, and simulate multi-step rolling forecasts; Set the prediction step size, calculate the prediction error and error amplification factor for each step, and determine the error accumulation phenomenon; The load standard deviation sequence is extracted to obtain the load fluctuation index. The error amplification factor is correlated with the load fluctuation index to construct an error propagation sensitivity network. Temperature, humidity, and wind speed parameters are extracted from meteorological data to calculate the daily temperature variation and the proportion of temperature-sensitive loads. Identify peak load times, statistically analyze peak and valley patterns on weekdays and non-weekdays, and calculate peak and valley fluctuation amplification factors; fit the error propagation coefficients for different date types to obtain error propagation correction factors; After normalizing the error propagation sensitivity, the proportion of temperature-sensitive loads, the peak-valley fluctuation amplification factor, and the error propagation correction factor, the comprehensive error propagation index is obtained through a weighted fusion formula.

3. The power demand forecasting method based on big data analysis according to claim 1, characterized in that, The specific operation steps of S3 are as follows: The preprocessed historical load sequence was decomposed into three levels using the Daubechies wavelet basis function to extract low-frequency trend components, mid-frequency periodic components, and high-frequency fluctuation components. The trend component is smoothed by moving average with a dynamic window length, and the predicted value of the trend component is obtained by least squares linear regression fitting. Frequency domain analysis of the periodic components is performed using fast Fourier transform to identify the dominant period and construct a period superposition model to obtain the predicted values ​​of the periodic components. Long Short-Term Memory (LSTM) networks are used to fit and predict short-term fluctuations in the fluctuation components. Once the predicted value of the fluctuation component exceeds the set short-term step size, it is set to zero. Based on the comprehensive error propagation index, a dynamic weight allocation strategy is designed to adaptively adjust with the prediction step size, and dynamic weight allocation is performed on each component to form a hierarchical independent prediction strategy.

4. The power demand forecasting method based on big data analysis according to claim 1, characterized in that, The specific operation steps of S4 are as follows: The predicted values ​​of trend component, cycle component and fluctuation component obtained from the hierarchical prediction are weighted and fused based on the corresponding dynamic weights to obtain the preliminary prediction value; An error estimation fitting network is constructed, with the actual error of the previous prediction step size as the input feature. A sliding window collects historical errors from multiple steps to form fitting samples. The time-series fitting is extrapolated through an autoregressive model to obtain the current step size error estimate. A confidence decay mechanism is introduced to calculate the confidence level corresponding to the prediction step size; The preliminary forecast and the baseline forecast are weighted and fused based on the forecast confidence level to obtain the compensated forecast; During the forecast execution process, the time points of the actual load data that have been acquired are monitored in real time, and the newly added actual load data are statistically analyzed point by point. When the cumulative actual load data of newly connected users reaches the preset time, a rolling update is triggered, and the error propagation-related parameters are recalculated and the weight coefficients and confidence decay coefficients of the autoregressive model are updated.

5. The power demand forecasting method based on big data analysis according to claim 1, characterized in that, The specific operation steps of S5 are as follows: The compensated predicted value sequence is divided into three validation segments—short-term, medium-term, and long-term—based on the prediction step size, and an accuracy threshold is set for each validation segment. Collect actual load data and calculate the prediction error index segment by segment; determine the prediction results based on the prediction error index of each verification segment and the corresponding set accuracy threshold. If the prediction result is deemed to have passed the verification, the multi-source parameters, including wavelet decomposition parameters, linear extrapolation equation coefficients, and error estimation fitting network parameters, are saved simultaneously to form a prediction configuration scheme. If the prediction result is determined to be locally optimized, then locate the current validation segment and adjust the error propagation correction factor and error estimation weight; If the prediction is determined to be unsuccessful, the configuration rollback mechanism is triggered to automatically restore the prediction parameters and configuration scheme that passed the previous round of verification, and an abnormal prediction log is generated. After verification, based on the deviation between the actual load value and the compensated predicted value, the exponential smoothing method is used to perform rolling correction on the subsequent prediction sequence. The correction weight decreases to zero as the prediction step size increases, and the load time series dataset is updated with the corrected prediction results for subsequent multi-step rolling prediction.

6. A system applied to the power demand forecasting method based on big data analysis as described in any one of claims 1-5, comprising: Data acquisition and preprocessing module: By acquiring historical power load, meteorological, date characteristics and power consumption structure data, and performing missing value imputation, abnormal data correction and standardization processing, a standardized load time series dataset is generated; Error propagation characteristics analysis module: Constructs a load time series matrix, calculates prediction error, error amplification factor and load fluctuation index, constructs an error propagation sensitivity network, and combines the proportion of temperature-sensitive loads and peak-valley fluctuation amplification factor to obtain a comprehensive error propagation index; Hierarchical prediction strategy generation module: Performs wavelet decomposition on the load sequence to predict the trend, periodic and fluctuation components respectively, and assigns dynamic weights based on the comprehensive error propagation index to form a hierarchical independent prediction strategy; Error Compensation Prediction Execution Module: Obtains preliminary prediction values ​​by weighted fusion of the prediction values ​​of each component, estimates the error by autoregressive model, and performs error compensation by combining confidence decay. Prediction result verification and correction module: Verify prediction accuracy by dividing the prediction into short, medium and long term, perform parameter saving, local optimization or configuration rollback based on prediction results, and use exponential smoothing rolling correction to update the load time series dataset.

Images