A method for predicting power load in industrial and commercial parks

Abstract

The present invention discloses a method for predicting the power of electric load in industrial and commercial parks, which relates to the technical field of electric load prediction, including obtaining the power data of electric equipment in the park, obtaining a standardized time series power matrix through normalization preprocessing; performing wavelet denoising and empirical mode decomposition on the power matrix to separate the power trend component and the fluctuation component; constructing a trend prediction sub-model using a long short-term memory network, and constructing a fluctuation prediction sub-model using a seasonal ARIMA model; weightedly combining the two sub-models through an integrated learning method to construct a comprehensive prediction model; finally, setting an adaptive threshold based on historical prediction errors for constraint processing, and outputting the final prediction result. The present invention adaptively separates power data features through a multi-level decomposition strategy, adopts a dual-model collaborative prediction mechanism, introduces environmental factors to improve prediction accuracy, and designs an adaptive correction mechanism to ensure the reliability of the results, thereby significantly improving the accuracy and stability of the prediction of the power load in the park.

Description

Technical Field

[0001] The invention relates to the technical field of electric load prediction, in particular to a method for predicting the power of electric load in an industrial and commercial park. Background Art

[0002] The electricity load forecasting of industrial and commercial parks is of great significance for the planning, operation and demand-side management of power systems. Currently, the commonly used forecasting methods mainly include time series method, regression analysis method and neural network method. These methods perform well when dealing with a single type of load change, but for scenarios such as parks where multiple types of buildings are mixed and load characteristics are complex, the forecasting accuracy is often unsatisfactory. Among them, traditional time series models are difficult to effectively capture long-term dependencies, neural network models are easily affected by noise, and have limitations when dealing with seasonal fluctuations with multiple cycles superimposed.

[0003] Especially in practical applications, the power load of the park is not only affected by its own operating rules, but also closely related to environmental factors such as temperature and humidity. Most existing methods separate these influencing factors and lack systematic modeling ideas, which makes it difficult to ensure the stability and reliability of the prediction results. At the same time, due to the randomness and diversity of building electrical equipment, load data often presents complex characteristics of superposition of trends and fluctuations, which brings huge challenges to the construction of prediction models. Summary of the invention

[0004] In view of the fact that the existing park power load prediction methods have the problems of low accuracy and poor generalization ability when dealing with complex load characteristics, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to construct a campus electricity load power prediction method that can effectively handle mixed scenes of multiple buildings, adapt to changes in environmental factors, and accurately predict trends and fluctuation characteristics.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides a method for predicting power load of an industrial and commercial park, which includes obtaining power data of power equipment in the park, marking timestamp marks and building identifications, and performing normalization preprocessing to obtain a standardized time series power matrix; performing wavelet denoising and empirical mode decomposition operations on the time series power matrix in sequence to separate power trend components and power fluctuation components; constructing a long short-term memory network, dividing the power trend component into a training sequence according to a time window, and using the Adam optimizer to optimize the network parameters to obtain a trend prediction sub-model; constructing a seasonal ARIMA model, using the power fluctuation component and environmental monitoring data as input variables, and training to obtain a fluctuation prediction sub-model; using an integrated learning method to weightedly combine the output results of the trend prediction sub-model and the fluctuation prediction sub-model to construct a comprehensive prediction model; sending the input data of the time period to be predicted into the comprehensive prediction model, setting an adaptive threshold based on the historical prediction error and performing constraint processing to obtain a final power prediction value sequence.

[0008] As a preferred solution of the power prediction method for the electric load of industrial and commercial parks described in the present invention, the method includes the following steps: performing wavelet denoising processing and empirical mode decomposition operation on the time series power matrix in sequence: performing wavelet decomposition operation, decomposing the time series power matrix into four layers through wavelet basis transformation, and obtaining a low-frequency power coefficient sequence and a high-frequency power coefficient sequence; calculating the corresponding noise threshold based on the mean and standard deviation of the high-frequency power coefficient sequence; performing soft threshold processing on the high-frequency power coefficient sequence according to the noise threshold to obtain a denoised high-frequency power coefficient sequence; performing wavelet reconstruction on the low-frequency power coefficient sequence and the denoised high-frequency power coefficient sequence to obtain a denoised power sequence; performing empirical mode decomposition on the denoised power sequence to extract an intrinsic mode function group and a residual sequence; dividing the intrinsic mode function group into a high-frequency group and a low-frequency group according to frequency characteristics, superimposing the high-frequency group as a power fluctuation component, and superimposing the low-frequency group and the residual sequence as a power trend component.

[0009] As a preferred solution of the power prediction method for the power load of industrial and commercial parks described in the present invention, the construction of a long short-term memory network includes the following steps: dividing the power trend component according to the preset time window length to generate a training sequence; constructing a long short-term memory network, the long short-term memory network includes an input layer, an output layer and multiple long short-term memory hidden layers; initializing the weight matrix and bias vector of the forgetting gate, input gate and output gate in the long short-term memory hidden layer; inputting the training sequence into the long short-term memory network, and calculating the hidden layer state sequence through the gating mechanism and the memory unit; mapping the hidden layer state sequence through the fully connected layer to obtain the predicted output, and calculating the mean square error function between the predicted output and the actual power trend value; using the Adam optimization algorithm to calculate the gradient of the mean square error function, iteratively updating the weight matrix and the bias vector until convergence to obtain a trend prediction sub-model.

[0010] As a preferred solution of the method for predicting power load of industrial and commercial parks described in the present invention, the seasonal ARIMA model is constructed, including the following steps: aligning the power fluctuation component with the environmental monitoring data according to the timestamp to generate a multidimensional joint sequence; performing differential operation and seasonal decomposition on the multidimensional joint sequence to eliminate the periodicity and trend to obtain a stationary data sequence; analyzing the stationary data sequence through the autocorrelation function and the partial autocorrelation function to obtain the ARIMA parameter group; setting the model structure parameters based on the ARIMA parameter group to establish a seasonal prediction equation containing seasonal cycle terms and environmental factor terms; using the maximum likelihood estimation method to calculate the autoregressive coefficient, moving average coefficient and seasonal term coefficient in the seasonal prediction equation to obtain the optimized model coefficient; using the white noise test and the Akaike information criterion to evaluate the fitting effect of the optimized model coefficient on the stationary data sequence, and outputting a fluctuation prediction sub-model that meets the preset accuracy threshold.

[0011] As a preferred solution of the method for predicting the power load of industrial and commercial parks described in the present invention, the construction of a comprehensive prediction model includes the following steps: predicting the trend prediction sub-model and the fluctuation prediction sub-model on the verification data set respectively to obtain a trend prediction value sequence and a fluctuation prediction value sequence; calculating the error evaluation index between the trend prediction sequence and the fluctuation prediction sequence and the actual value in the verification data set; constructing a combined weight optimization objective function based on the error evaluation index; performing a particle swarm optimization operation to iteratively solve the combined weight optimization objective function to obtain an optimal combined weight vector; performing a weighted combination operation to weightedly combine the trend prediction sub-model and the fluctuation prediction sub-model according to the optimal combined weight vector to construct a comprehensive prediction model.

[0012] As a preferred solution of the power prediction method for the power load of the industrial and commercial park described in the present invention, the method includes the following steps: setting an adaptive threshold based on the historical prediction error and performing constraint processing: calculating the power change of adjacent sampling points in the historical power data, sorting the power change in ascending order, and obtaining a power change sequence; selecting a Gaussian kernel function to construct a kernel density estimator, performing density fitting on the power change sequence, and obtaining the mean and standard deviation of the probability density function; setting a bilateral constraint interval according to the mean and the standard deviation to obtain the upper and lower limit thresholds; obtaining the power prediction value sequence output by the comprehensive prediction model, calculating the difference between the power prediction values ​​at adjacent moments, comparing it with the bilateral constraint interval, and recording the time index of the exceeding point; for each exceeding point, extracting the historical power data of the same period corresponding to the time index, setting an exponential decay coefficient, and performing a time-weighted average operation to output a corrected reference value; for each exceeding point, calculating the deviation between the power difference and the center of the constraint interval, setting a linear combination weight based on the deviation, and weighted combining the corrected reference value with the corresponding predicted power value to obtain a corrected power prediction value sequence.

[0013] As a preferred scheme of the method for predicting the power load of industrial and commercial parks described in the present invention, setting the bilateral constraint interval according to the mean and standard deviation includes: taking the mean as the central value of the bilateral constraint interval; calculating the multiple product of the standard deviation and the confidence interval coefficient; adding and subtracting the multiple product of the central value to obtain the upper and lower limit thresholds of the bilateral constraint interval.

[0014] As a preferred solution of the method for predicting the power load of industrial and commercial parks described in the present invention, the ARIMA parameter group includes the order of autoregressive terms, the order of differential terms and the order of moving average terms.

[0015] The beneficial effects of the present invention are as follows: first, the present invention uses an improved empirical mode decomposition method to adaptively separate the trend and fluctuation components of power data, avoiding the decomposition deviation caused by improper selection of basis functions in traditional methods; secondly, the trend component is modeled using a long short-term memory network, and the ability to capture long-term dependencies is improved through gating mechanisms and optimization strategies; the power fluctuation component is modeled using an improved ARIMA model, and the adaptability of the model to changes in external conditions is enhanced by introducing environmental factors and seasonal cycle terms; finally, the optimal combination of sub-models is achieved through an integrated learning method based on particle swarm optimization, and combined with adaptive threshold constraints and a correction mechanism based on historical data, the temporal smoothness and accuracy of the prediction results are ensured. This multi-level, multi-model collaborative prediction scheme effectively improves the overall performance of the park power load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 The overall flow chart of the power prediction method for the power load in industrial and commercial parks.

[0018] Figure 2 The power data decomposition flow chart of the power prediction method for the electrical load in industrial and commercial parks.

[0019] Figure 3 Construct a flow chart for the prediction model of the power prediction method of the electrical load in the industrial and commercial park. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1~Figure 3 , which is the first embodiment of the present invention, and provides a method for predicting the power load of an industrial and commercial park. The comprehensive prediction model is as follows Figure 1 As shown, including,

[0024] S1: Obtain the power data of the electrical equipment in the park, mark the timestamp and building identification, and perform normalization preprocessing to obtain a standardized time series power matrix.

[0025] Specifically, the original sampled power of electrical equipment in each building is collected through the park energy consumption monitoring system, including the electrical equipment number, sampling time and corresponding power value, and the sampling time interval is 15 minutes; the collected original sampled power is classified and encoded according to the building type, and the building identification field is generated. The original sampled power under the same building identification is summarized to obtain the total power of the building; the total power of the building is divided into several time segments with a 24-hour cycle, and each time segment is marked with a timestamp in a unified format; the total power of the building with timestamp marks and building identifications is arranged in time series to construct an original power data matrix, in which the rows of the original power data matrix represent different times and the columns represent different buildings; the power data in the original power data matrix is ​​standardized by the maximum and minimum normalization method; the standardized power data is integrated with the corresponding timestamp marks and building identifications to generate a standardized time series power matrix.

[0026] It should be noted that the standardized time series power matrix contains three dimensions: time dimension, building dimension and power value dimension. The time dimension records the specific time of data sampling, the building dimension identifies the type and number of different buildings, and the power value dimension stores the normalized power data.

[0027] S2: Perform wavelet denoising and empirical mode decomposition operations on the time series power matrix in sequence to separate the power trend component and the power fluctuation component.

[0028] Specifically, the power data decomposition flow chart is as follows: Figure 2 As shown, the following steps are included:

[0029] S2.1: Execute wavelet decomposition operation, decompose the time series power matrix into four layers through wavelet basis transformation, and obtain low-frequency power coefficient sequence and high-frequency power coefficient sequence.

[0030] The db4 wavelet basis function is used for wavelet basis transformation, which has good orthogonality and compact support, and helps to improve the accuracy of signal decomposition. The selection of the four-layer decomposition scale is based on the sampling frequency of the power data and the characteristics of the target frequency band, which can effectively separate different frequency components.

[0031] S2.2: Based on the mean and standard deviation of the high-frequency power coefficient sequence, the corresponding noise threshold is calculated.

[0032] Specifically, the noise threshold is calculated using the VisuShrink method. The threshold size is related to the standard deviation and sample size of the high-frequency coefficient sequence, and can adaptively adjust the denoising intensity.

[0033] S2.3: Perform soft threshold processing on the high-frequency power coefficient sequence according to the noise threshold to obtain a denoised high-frequency power coefficient sequence.

[0034] Among them, the soft threshold processing adopts the wavelet coefficient shrinkage method to maintain the continuity of the signal through smoothing processing and avoid additional oscillations.

[0035] S2.4: Perform wavelet reconstruction on the low-frequency power coefficient sequence and the denoised high-frequency power coefficient sequence to obtain the denoised power sequence.

[0036] It should be noted that wavelet reconstruction uses a multi-resolution analysis method to ensure the integrity and accuracy of the reconstructed signal.

[0037] S2.5: Perform empirical mode decomposition on the denoised power sequence to extract the intrinsic mode function group and the residual sequence.

[0038] In the specific implementation, all local maximum points and local minimum points in the denoised power sequence are identified; the maximum points and the minimum points are connected by cubic spline interpolation to obtain the upper envelope and the lower envelope; the mean of the upper envelope and the lower envelope is calculated to obtain the mean envelope; the mean envelope is subtracted from the denoised power sequence to obtain a candidate intrinsic mode function; it is determined whether the candidate intrinsic mode function meets the definition conditions of the intrinsic mode function: the difference between the number of local extreme points and the number of zero-crossing points does not exceed a preset value, and the mean of the upper and lower envelopes is approximately zero; if the conditions are not met, the candidate intrinsic mode function is used as a new input sequence, and the above decomposition process is repeated until the definition conditions are met to obtain the first intrinsic mode function; the obtained intrinsic mode function is subtracted from the denoised power sequence, and the above decomposition process is repeated for the remaining sequence until the remaining sequence becomes a monotonic function or is lower than a preset threshold, and the remaining sequence is the residual sequence.

[0039] S2.6: Divide the intrinsic mode function group into a high-frequency group and a low-frequency group according to the frequency characteristics, superimpose the high-frequency group into a power fluctuation component, and superimpose the low-frequency group and the residual sequence into a power trend component.

[0040] Furthermore, the average period of each intrinsic mode function is calculated, and the average period is determined by the average time interval between adjacent extreme points; the period threshold is set to 24 hours, and the intrinsic mode functions with an average period less than the threshold are classified into the high-frequency group, and the intrinsic mode functions with an average period greater than or equal to the threshold are classified into the low-frequency group; the intrinsic mode functions in the high-frequency group are superimposed to obtain the power fluctuation component that characterizes the short-term fluctuation characteristics; the intrinsic mode functions in the low-frequency group are superimposed with the residual sequence to obtain the power trend component that characterizes the long-term change trend.

[0041] The beneficial effect is that, through the above improved empirical mode decomposition method, the present invention can adaptively separate the trend component and the fluctuation component in the power data without presetting the decomposition basis function, thus avoiding the decomposition deviation caused by improper selection of basis functions in the traditional method. At the same time, the component division method based on periodic characteristics fully considers the periodic characteristics of the power load data, and improves the accuracy and physical significance of the component division.

[0042] S3: Construct a long short-term memory network, divide the power trend component into training sequences according to the time window, and use the Adam optimizer to optimize the network parameters to obtain the trend prediction sub-model.

[0043] Specifically, the following steps are included:

[0044] S3.1: Divide the power trend component according to the preset time window length to generate a training sequence.

[0045] S3.2: Construct a long short-term memory network, which includes an input layer, an output layer, and multiple long short-term memory hidden layers.

[0046] S3.3: Initialize the weight matrices and bias vectors of the forget gate, input gate, and output gate in the LSTM hidden layer.

[0047] S3.4: Input the training sequence into the long short-term memory network, and calculate the hidden state sequence through the gating mechanism and memory units.

[0048] In specific implementation, the forget gate receives the current input and the previous hidden state , calculate the forgetting coefficient through the sigmoid function ; The input gate also calculates the importance of the current input information and candidate memory cell states ; Update memory cell status ,in Represents element-by-element multiplication; the output gate generates the hidden state at the current moment according to the updated memory unit state The gating mechanism can effectively capture long-term dependencies by adaptively regulating the flow of information.

[0049] S3.5: Map the hidden layer state sequence through the fully connected layer to obtain the predicted output, and calculate the mean square error function between the predicted output and the actual power trend value.

[0050] In order to improve the robustness of the model, the fully connected layer includes a dropout mechanism with a dropout rate of 0.2, which reduces the risk of model overfitting by randomly inactivating some neuron connections. The loss function uses the mean square error with an L2 regularization term, and the regularization coefficient is 0.001.

[0051] S3.6: Use the Adam optimization algorithm to calculate the gradient of the mean square error function, and iteratively update the weight matrix and bias vector until convergence to obtain the trend prediction sub-model.

[0052] In the optimization phase, the initial learning rate of the Adam optimizer is set to 0.001, and the momentum decay parameter is 0.9, the second-order momentum decay parameter is 0.999; a small batch stochastic gradient descent method with a batch_size of 32 is used; a learning rate adaptive adjustment strategy is designed. When the validation set loss has not improved for 5 consecutive epochs, the learning rate is reduced to 0.1 times the original value; when the validation set loss has not improved for 10 consecutive epochs, the training is stopped and the optimal model parameters are saved.

[0053] S4: Construct a seasonal ARIMA model, take the power fluctuation component and environmental monitoring data as input variables, and train to obtain the fluctuation prediction sub-model.

[0054] Specifically, the following steps are included:

[0055] S4.1: Align the power fluctuation component with the environmental monitoring data by timestamp to generate a multi-dimensional joint sequence.

[0056] Among them, the multidimensional joint sequence includes temperature features, humidity features, and wind speed features.

[0057] S4.2: Perform difference operation and seasonal decomposition on the multidimensional joint series to eliminate periodicity and trend to obtain a stationary data series.

[0058] Specifically, the difference operation uses the forward difference method to eliminate long-term trends, and the seasonal decomposition is based on the X-12-ARIMA method to extract the cyclical components, which can effectively identify and remove multiple seasonal patterns.

[0059] S4.3: Analyze the stationary data series through the autocorrelation function and partial autocorrelation function to obtain the ARIMA parameter group.

[0060] Among them, the ARIMA parameter group includes the order of autoregressive term, the order of difference term and the order of moving average term.

[0061] S4.4: Set the model structure parameters based on the ARIMA parameter group and establish a seasonal prediction equation containing seasonal cycle terms and environmental factor terms.

[0062] In specific implementation, the construction process of the prediction equation is as follows: construct an autoregressive term expression, express the power value at the current moment as a linear combination of the power values ​​at the previous p moments, and obtain an autoregressive polynomial; construct a differential term expression, perform a d-order differential operation on the input sequence, and obtain a differential term processing expression; construct a moving average term expression, express the prediction error at the current moment as a linear combination of the prediction errors at the previous q moments, and obtain a moving average polynomial; construct a seasonal cycle term expression, identify the three main cycles of 24 hours, 7 days and 30 days based on historical data, and establish corresponding cycle term expressions respectively; construct an environmental factor influence term expression, map the temperature, humidity and wind speed characteristics into linear influence terms, and introduce nonlinear interaction terms to capture the coupling relationship between environmental factors; combine the above five expressions according to a modular structure to obtain a complete prediction equation.

[0063] S4.5: Use the maximum likelihood estimation method to calculate the autoregressive coefficient, moving average coefficient and seasonal term coefficient in the seasonal prediction equation to obtain the optimized model coefficient.

[0064] In specific implementation, the model coefficient optimization process is as follows: initialize the model coefficients, use the Yule-Walker equation to calculate the initial value of the autoregressive coefficient, and use the least squares method to estimate the initial value of the moving average coefficient and the seasonal term coefficient; construct the log-likelihood function, substitute the prediction equation into the likelihood function expression, and establish the optimization objective function; set parameter constraints to ensure the stationarity of the autoregressive model and the reversibility of the moving average model; use the quasi-Newton method to iteratively solve the optimal coefficients, and update the Hessian matrix approximation in each iteration; stop the iteration when the coefficient change between two adjacent iterations is less than the preset threshold, and output the optimal model coefficients.

[0065] S4.6: Use white noise test and Akaike information criterion to evaluate the fitting effect of the optimization model coefficients on the stationary data series, and output the volatility prediction sub-model that meets the preset accuracy threshold.

[0066] In specific implementation, the model evaluation process is as follows: calculate the model residual sequence, and take the difference between the actual observation value and the model predicted value as the residual sample; perform the Ljung-Box test, calculate the autocorrelation coefficient of the residual sequence, and verify whether it obeys the white noise characteristics; count the number of model parameters, calculate the data sample size, and substitute it into the Akaike Information Criterion formula to obtain the AIC value; set the significance level of the white noise test to 0.05, and require the p value to be greater than the significance level; compare the calculated AIC value with the preset threshold to ensure that the model complexity is within a reasonable range; when the model meets the requirements of both the white noise test and the AIC criterion, it is determined as the final volatility prediction sub-model.

[0067] The beneficial effect is that the present invention realizes accurate modeling of power fluctuation components through an improved ARIMA model construction method. The introduction of environmental factors enhances the adaptability of the model to changes in external conditions, and the design of seasonal cycle terms improves the ability to capture periodic fluctuations. At the same time, the model evaluation mechanism based on statistical tests ensures the reliability and stability of the prediction results.

[0068] S5: Using the ensemble learning method, the output results of the trend prediction sub-model and the fluctuation prediction sub-model are weighted and combined to construct a comprehensive prediction model.

[0069] Specifically, the comprehensive prediction model construction flow chart is as follows: Figure 3 As shown, the following steps are included:

[0070] S5.1: The trend prediction sub-model and the fluctuation prediction sub-model are predicted on the validation data set respectively to obtain the trend prediction value sequence and the fluctuation prediction value sequence.

[0071] S5.2: Calculate the error evaluation index between the trend forecast series and the volatility forecast series and the actual values ​​in the validation data set.

[0072] Among them, the error evaluation indicators include mean absolute percentage error MAPE, root mean square error RMSE, and mean absolute error MAE.

[0073] S5.3: Construct a combined weight optimization objective function based on the error evaluation index. The objective function takes the form of the weighted sum of squares of the prediction error.

[0074] Specifically, firstly, the three error evaluation indicators of MAPE, RMSE and MAE of the trend prediction value sequence are obtained, and the ratio of each indicator to the mean of the validation data set is calculated to obtain a normalized error indicator group; secondly, the three error evaluation indicators of MAPE, RMSE and MAE of the fluctuation prediction value sequence are obtained, and the ratio of each indicator to the mean of the validation data set is calculated to obtain a normalized error indicator group; then, a combined weight optimization objective function is constructed, based on the normalized error indicator group of the trend prediction value sequence and the fluctuation prediction value sequence, and the weighted square sum form of the prediction error is adopted; finally, the combined weight optimization objective function is mathematically derived, and the gradient calculation expression of the objective function is established.

[0075] S5.4: Execute particle swarm optimization operation, iteratively solve the combination weight optimization objective function, and obtain the optimal combination weight vector.

[0076] Specifically, first, the initial parameters of the particle swarm are set, including the number of particles, the number of iterations and the convergence condition, and the position and velocity of each particle are initialized; secondly, the position of each particle is substituted into the combined weight optimization objective function, the fitness value is calculated, and the individual optimal position and the global optimal position are recorded; then, the particle velocity is updated according to the gradient calculation expression of the objective function, and the particle position is adjusted based on the updated velocity; then, the fitness calculation and position update operations are repeated until the preset iteration termination condition is reached; finally, the global optimal position at convergence is determined as the optimal combined weight vector.

[0077] S5.5: Perform a weighted combination operation to weight the trend prediction sub-model and the fluctuation prediction sub-model according to the optimal combination weight vector to construct a comprehensive prediction model.

[0078] S6: The input data of the time period to be predicted is sent to the comprehensive prediction model, an adaptive threshold is set based on the historical prediction error, and constraint processing is performed to obtain the final power prediction value sequence.

[0079] Specifically, the following steps are included:

[0080] S6.1: Calculate the power changes of adjacent sampling points in the historical power data, sort the power changes in ascending order, and obtain a power change sequence.

[0081] S6.2: Select the Gaussian kernel function to construct a kernel density estimator, perform density fitting on the power change series, and obtain the mean and standard deviation of the probability density function.

[0082] Specifically, the Gaussian kernel function has good smoothness and continuity, and can accurately characterize the distribution characteristics of the power change sequence. The probability density function obtained by kernel density estimation reflects the statistical law of power change.

[0083] S6.3: Set the bilateral constraint interval based on the mean and standard deviation to obtain the upper and lower limit thresholds.

[0084] During the implementation process, the mean is used as the central value of the bilateral constraint interval; the multiple product of the standard deviation and the confidence interval coefficient is calculated; the central value is added or subtracted from the multiple product to obtain the upper and lower limit thresholds of the bilateral constraint interval. The confidence interval coefficient is determined based on the distribution characteristics of the power change sequence. By calculating the data coverage of the power change sequence under different standard deviation multiples, the standard deviation multiple corresponding to the data coverage reaching 95% is selected as the confidence interval coefficient. This method not only ensures the effective identification of abnormal fluctuations by the constraint interval, but also avoids the reasonable changes of the prediction results affected by overly strict constraints.

[0085] S6.4: Obtain the power prediction value sequence output by the comprehensive prediction model, calculate the difference between the power prediction values ​​at adjacent moments, compare it with the bilateral constraint interval, and record the time index of the exceeding point.

[0086] S6.5: For each exceeding limit point, extract the historical power data of the same period corresponding to the time index, set the exponential decay coefficient, perform time-weighted average operation and output the corrected reference value.

[0087] Specifically, for each over-limit point, the power data of three hours before and after the corresponding moment in the historical data is extracted according to its time index to construct a historical power data set; trace back seven days from the current prediction moment, and extract the power data sequence of the corresponding time period in the historical power data set within these seven days; set the exponential decay coefficient to 0.8, and calculate the time weight sequence based on this coefficient, so that the historical data closer to the current prediction moment obtains a larger weight value; perform weighted averaging operation on the power data sequence of the corresponding time period of each day and the corresponding time weight to obtain the corrected reference value.

[0088] S6.6: For each exceeding limit point, calculate the deviation between the power difference and the center of the constraint interval, set the linear combination weight based on the deviation, perform weighted combination of the corrected reference value and the corresponding predicted power value, and obtain a corrected power prediction value sequence.

[0089] Specifically, the difference between the power prediction value of the out-of-limit point and the center value of the constraint interval is calculated, and the difference is divided by the half-width of the constraint interval to obtain the normalized deviation; the linear combination weight is set based on the normalized deviation, and when the deviation is larger, the weight of the corrected reference value is larger, and the weight of the predicted power value is smaller, and vice versa; according to the proportional relationship between the weight of the corrected reference value and the weight of the predicted power value, the two are linearly combined to obtain the corrected power prediction value; all out-of-limit points are corrected according to the above method to obtain the final power prediction value sequence.

[0090] Preferably, by considering historical data of the same period and introducing time decay weights, the correction process better maintains the time series characteristics of load changes; at the same time, the adaptive weight allocation strategy based on deviation not only ensures the effective correction of abnormal prediction values, but also avoids the loss of prediction accuracy caused by over-correction. This method can significantly improve the time series smoothness and accuracy of the prediction results.

[0091] In summary, the present invention firstly uses the improved empirical mode decomposition method to adaptively separate the trend and fluctuation components of power data, avoiding the decomposition deviation caused by improper selection of basis functions in traditional methods; secondly, the trend component is modeled using the long short-term memory network, and the ability to capture long-term dependencies is improved through the gating mechanism and optimization strategy; the power fluctuation component is modeled using the improved ARIMA model, and the adaptability of the model to changes in external conditions is enhanced by introducing environmental factors and seasonal cycle terms; finally, the optimal combination of sub-models is achieved through an integrated learning method based on particle swarm optimization, and combined with adaptive threshold constraints and a correction mechanism based on historical data, the time series smoothness and accuracy of the prediction results are ensured. This multi-level, multi-model collaborative prediction scheme effectively improves the overall performance of the park power load prediction.

[0092] Example 2, reference Figure 1~Figure 3 , which is the second embodiment of the present invention, and this embodiment provides a method for predicting the power of electricity load in an industrial and commercial park. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0093] To verify the effectiveness of the present invention, the power load data of an industrial park from January to December 2023 was selected for experimental verification. The park contains 12 buildings, including 4 office buildings, 5 production plants, and 3 storage buildings, with a sampling interval of 15 minutes. The experimental environment is an Intel Core i9-13900K processor, 128GB memory, and is implemented using Python 3.9 programming.

[0094] First, the raw data was preprocessed. The total number of samples collected by the park energy consumption monitoring system was 35,040 (one sampling point every 15 minutes, a total of 35,040 time points throughout the year). The collected data was classified and coded according to the building type to generate a standardized time series power matrix. After data preprocessing, 80% of the data was used as a training set, 10% as a validation set, and the remaining 10% as a test set.

[0095] In the signal decomposition stage, the db4 wavelet basis function is used for four-layer decomposition. The low-frequency coefficient sequence obtained after wavelet transform reflects the main change trend of the load, and the high-frequency coefficient sequence contains short-term fluctuation information. The noise threshold calculated by the VisuShrink method is about 0.15. After soft threshold processing of the high-frequency coefficient, the signal-to-noise ratio is increased from the original 15.3dB to 21.8dB. The empirical mode decomposition obtains 8 intrinsic mode functions (IMFs), among which the average period of IMF1-IMF3 is less than 24 hours, which is classified into the high-frequency group to constitute the power fluctuation component, and the remaining IMFs and residual terms constitute the power trend component.

[0096] When constructing the trend prediction sub-model, the LSTM network adopts a 3-layer structure, each layer contains 128 neurons, and the dropout rate is set to 0.2. During the training process, batch_size is set to 32, the initial learning rate is 0.001, and when the validation set loss has not improved for 5 consecutive epochs, the learning rate is reduced to 0.1 times the original. After about 200 iterations, the loss function of the model on the validation set tends to be stable, and MAPE drops to 3.2%.

[0097] In the volatility prediction submodel, the parameters of the seasonal ARIMA model were determined by grid search, and the ARIMA(2,1,2)(1,1,1)24 model was finally selected. Environmental monitoring data included temperature (15-35℃), relative humidity (45%-85%), and wind speed (0-8m / s). The MAPE of the model on the validation set was 4.1%, and the Ljung-Box test (p value was 0.087>0.05) and AIC criterion evaluation (AIC value was -3245.6) showed that the model had a good fitting effect.

[0098] In order to evaluate the performance of the method of the present invention, it is compared with the traditional LSTM model, ARIMA model and BP neural network. The experimental results are shown in Table 1.

[0099] Table 1 Performance comparison of different prediction methods

[0100] Prediction Methods MAPE(%) RMSE(kW) MAE(kW) Computation time (s) LSTM 5.8 245.3 198.6 156.4 ARIMA 6.2 267.8 215.3 89.2 BP Neural Network 7.1 289.4 234.7 102.8 The present invention 3.4 156.7 127.9 198.5

[0101] As can be seen from Table 1, the method of the present invention is superior to the comparative method in all evaluation indicators. In particular, in terms of the MAPE indicator, it is improved by 41.4% compared with the traditional LSTM model, 45.2% compared with the ARIMA model, and 52.1% compared with the BP neural network. Although the calculation time of the present invention is slightly longer, considering the significant improvement in prediction accuracy, this calculation overhead is acceptable.

[0102] Finally, in the prediction result optimization stage, the power change probability density function obtained by kernel density estimation presents an approximate normal distribution feature, with a mean of 0.142kW and a standard deviation of 2.876kW. A 95% confidence interval was set, and the upper and lower thresholds of the power change were obtained to be ±5.63kW respectively. For the prediction points exceeding the threshold, an exponential decay coefficient of 0.8 was used to perform weighted correction on the historical data. Finally, the prediction result MAPE on the test set was reduced to 3.4%, the RMSE was 156.7kW, and the MAE was 127.9kW, which fully verified the effectiveness and practicality of the method of the present invention.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting power load in an industrial and commercial park, characterized by: include, Obtain the power data of the park's electrical equipment, mark the timestamp and building identification, and perform normalization preprocessing to obtain a standardized time series power matrix; Performing wavelet denoising and empirical mode decomposition operations on the time series power matrix in sequence to separate a power trend component and a power fluctuation component; A long short-term memory network is constructed, the power trend component is divided into training sequences according to time windows, and the network parameters are optimized using the Adam optimizer to obtain a trend prediction sub-model; A seasonal ARIMA model is constructed, the power fluctuation component and the environmental monitoring data are used as input variables, and a fluctuation prediction sub-model is obtained by training; An integrated learning method is used to weight and combine the output results of the trend prediction sub-model and the fluctuation prediction sub-model to construct a comprehensive prediction model; The input data of the time period to be predicted is sent to the comprehensive prediction model, and an adaptive threshold is set based on the historical prediction error and constraint processing is performed to obtain a final power prediction value sequence.

2. The method for predicting power load of an industrial and commercial park according to claim 1, characterized in that: The step of sequentially performing wavelet denoising and empirical mode decomposition operations on the time series power matrix comprises the following steps: Perform wavelet decomposition operation, decompose the time series power matrix into four layers through wavelet basis transformation, and obtain low-frequency power coefficient sequence and high-frequency power coefficient sequence; Based on the mean and standard deviation of the high-frequency power coefficient sequence, a corresponding noise threshold is calculated; Performing soft threshold processing on the high-frequency power coefficient sequence according to the noise threshold to obtain a high-frequency power coefficient sequence after noise reduction; Performing wavelet reconstruction on the low-frequency power coefficient sequence and the denoised high-frequency power coefficient sequence to obtain a denoised power sequence; Performing empirical mode decomposition on the noise reduction power sequence to extract an intrinsic mode function group and a residual sequence; The intrinsic mode function group is divided into a high-frequency group and a low-frequency group according to frequency characteristics, the high-frequency group is superimposed as a power fluctuation component, and the low-frequency group and the residual sequence are superimposed as a power trend component.

3. The method for predicting power load of an industrial and commercial park according to claim 1, characterized in that: The construction of the long short-term memory network comprises the following steps: Dividing the power trend component according to a preset time window length to generate a training sequence; Constructing a long short-term memory network, wherein the long short-term memory network includes an input layer, an output layer, and a plurality of long short-term memory hidden layers; Initializing the weight matrix and bias vector of the forget gate, input gate and output gate in the long short-term memory hidden layer; Input the training sequence into the long short-term memory network, and calculate the hidden layer state sequence through the gating mechanism and the memory unit; Mapping the hidden layer state sequence through a fully connected layer to obtain a predicted output, and calculating a mean square error function between the predicted output and an actual power trend value; The Adam optimization algorithm is used to calculate the gradient of the mean square error function, and the weight matrix and the bias vector are iteratively updated until convergence to obtain a trend prediction sub-model.

4. The method for predicting power load of an industrial and commercial park according to claim 1, characterized in that: The construction of seasonal ARIMA model includes the following steps: The power fluctuation component is aligned with the environmental monitoring data by timestamp to generate a multi-dimensional joint sequence; Performing difference operation and seasonal decomposition on the multi-dimensional joint sequence to eliminate periodicity and trend to obtain a stationary data sequence; Analyze the stationary data sequence through autocorrelation function and partial autocorrelation function to obtain an ARIMA parameter group; Setting model structure parameters based on the ARIMA parameter group to establish a seasonal prediction equation including seasonal cycle terms and environmental factor terms; The maximum likelihood estimation method is used to calculate the autoregressive coefficient, the moving average coefficient and the seasonal term coefficient in the seasonal prediction equation to obtain the optimized model coefficient; The white noise test and Akaike information criterion are used to evaluate the fitting effect of the optimization model coefficients on the stationary data sequence, and a fluctuation prediction sub-model that meets a preset accuracy threshold is output.

5. The method for predicting power load of an industrial and commercial park according to claim 1, characterized in that: The construction of the comprehensive prediction model comprises the following steps: The trend prediction sub-model and the fluctuation prediction sub-model are respectively predicted on the validation data set to obtain a trend prediction value sequence and a fluctuation prediction value sequence; Calculate the error evaluation index between the trend prediction sequence and the fluctuation prediction sequence and the actual value in the validation data set; Constructing a combined weight optimization objective function based on the error evaluation index; Execute a particle swarm optimization operation to iteratively solve the combination weight optimization objective function to obtain an optimal combination weight vector; A weighted combination operation is performed to weight the trend prediction sub-model and the fluctuation prediction sub-model according to the optimal combination weight vector to construct a comprehensive prediction model.

6. The method for predicting power load of an industrial and commercial park according to claim 1, characterized in that: The method of setting an adaptive threshold based on historical prediction errors and performing constraint processing comprises the following steps: Calculating power changes of adjacent sampling points in the historical power data, and sorting the power changes in ascending order to obtain a power change sequence; Selecting a Gaussian kernel function to construct a kernel density estimator, performing density fitting on the power change sequence, and obtaining a mean and a standard deviation of a probability density function; Setting a bilateral constraint interval according to the mean and standard deviation to obtain upper and lower limit thresholds; Obtaining a power prediction value sequence output by a comprehensive prediction model, calculating a difference between power prediction values ​​at adjacent moments, comparing the difference with the bilateral constraint interval, and recording a time index of a point where the limit is exceeded; For each exceeding limit point, extract the historical power data of the same period corresponding to the time index, set the exponential decay coefficient, perform time-weighted average operation and output the correction reference value; For each exceeding limit point, the deviation between the power difference and the center of the constraint interval is calculated, a linear combination weight is set based on the deviation, the revised reference value and the corresponding predicted power value are weighted combined to obtain a revised power prediction value sequence.

7. The method for predicting power load of an industrial and commercial park according to claim 6, characterized in that: Setting a bilateral constraint interval based on the mean and standard deviation includes: The mean is taken as the center value of the bilateral constraint interval; Calculate the product of the standard deviation and the confidence interval coefficient; The center value is added or subtracted from the multiple product to obtain the upper and lower limit thresholds of the bilateral constraint interval.

8. The method for predicting power load of an industrial and commercial park according to claim 4, characterized in that: The ARIMA parameter group includes the autoregressive term order, the difference term order and the moving average term order.

Images