An industrial park load forecasting method and system fusing production features and multi-granularity timing

Abstract

This invention discloses a method and system for load forecasting in industrial parks that integrates production characteristics and multi-granularity time series data in the field of power system load forecasting technology. The method includes: preprocessing historical load data for the industrial park over a specific time period, as well as meteorological and production day data for a target time period, to obtain processed historical load data and meteorological and production day data for the target time period; constructing a production characteristic set for the target time period based on the processed meteorological and production day data; constructing a multi-granularity load sequence for the historical time period based on the processed historical load data; and generating a predicted load sequence for the target time period using a trained industrial park load forecasting model based on the multi-granularity load sequence for the historical time period and the production characteristic set for the target time period. This invention is applicable to power management scenarios in industrial parks with complex production scheduling patterns and has high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a method and system for forecasting industrial park loads that integrates production characteristics and multi-granularity time series. Background Technology

[0002] As an important carrier of regional economy, industrial parks have significant unique characteristics in their electricity load. Unlike residential and commercial loads, industrial park loads are affected by various factors such as production plans, shift schedules, and equipment maintenance, exhibiting nonlinear and multi-timescale coupling characteristics. Accurate load forecasting is of great significance for park power dispatch, demand-side management, and energy optimization.

[0003] Existing load forecasting methods mainly include traditional statistical methods, machine learning methods, and deep learning methods. Traditional statistical methods, such as time series analysis and regression analysis, have certain limitations in handling nonlinear load characteristics. Machine learning methods, such as support vector machines and random forests, can handle nonlinear relationships, but their ability to model time-series dependencies is relatively limited. Deep learning methods, such as LSTM, perform well in time-series modeling, but they are mostly designed for general power load scenarios and do not adequately consider the production characteristics of industrial parks.

[0004] In the practical application of load forecasting in industrial parks, the existing technologies have the following areas for improvement: (1) Most methods use general load characteristics and do not make full use of the production characteristics unique to industrial parks, such as shift codes, cumulative production time, and production saturation; (2) Most methods use a unified weight design loss function and need to strengthen the differentiated processing of key periods such as production peaks and regular operation; (3) Existing methods mostly use fixed granularity modeling and there are few schemes that combine the production day type of industrial parks to construct dynamic samples; (4) The post-processing correction of existing technologies mostly relies on a single meteorological factor, and the application of adaptive correction mechanisms based on historical error patterns and physical constraint verification is relatively limited.

[0005] Therefore, it is necessary to provide a load forecasting method that is tailored to the characteristics of industrial park scenarios, integrates multi-source data, has multi-granularity modeling capabilities, and an adaptive correction mechanism, in order to improve forecasting accuracy and practicality. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems in the existing technology, such as the industrial park load being significantly affected by production plans, the inability to effectively integrate multi-source data, and the lack of prediction optimization for critical periods. It provides an industrial park load prediction method and system that integrates production characteristics and multi-granularity time series. By introducing multi-source data consistency verification, production characteristic construction, multi-granularity time series modeling, multi-task weighted loss function, multi-task learning framework, and adaptive correction mechanism, the prediction accuracy and practicality are improved.

[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0008] In a first aspect, the present invention provides a method for predicting the load of industrial parks that integrates production characteristics and multi-granularity time series, comprising:

[0009] Data preprocessing is performed on the historical load data of the industrial park, as well as the meteorological data and production day data for the target time period, to obtain the processed historical load data, as well as the meteorological data and production day data for the target time period.

[0010] Based on the processed meteorological data and production day data for the target time period, a production feature set for the target time period is constructed.

[0011] Construct a multi-granularity load sequence for the historical time period based on the processed historical time period load data;

[0012] Based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period, the predicted load sequence of the target time period is generated using the trained industrial park load prediction model.

[0013] Secondly, the present invention provides an industrial park load forecasting system that integrates production characteristics and multi-granularity time series, comprising:

[0014] The data preprocessing module is used to preprocess the load data of the industrial park for historical time periods, as well as the meteorological data and production day data for the target time period, to obtain the processed load data of the historical time periods, as well as the meteorological data and production day data for the target time period.

[0015] The production feature construction module is used to: construct a production feature set for the target time period based on the processed meteorological data and daily production data for the target time period;

[0016] A multi-granularity load sequence construction module is used to: construct a multi-granularity load sequence for a historical time period based on the processed historical time period load data;

[0017] The load forecasting module is used to generate a predicted load sequence for the target time period based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period using a trained industrial park load forecasting model.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0019] 1. The proposed industrial park load forecasting method integrates production characteristics and multi-granularity time series. It introduces unique industrial park features such as daily type coding, shift characteristics, cumulative production duration, and production saturation to effectively capture the impact of production plans on load. It constructs three-level time series (fine-grained, medium-grained, and coarse-grained) and, through attention mechanism fusion, can simultaneously capture minute-level fluctuations, hourly patterns, and daytime trends. Differentiated weights are set for peak production periods, regular operation periods, and off-peak periods to prioritize forecast accuracy during critical periods. Combined with meteorological correction, historical error correction, and physical constraint verification, the method ensures accurate and reasonable forecast results.

[0020] 2. The industrial park load forecasting system proposed in this invention integrates production characteristics and multi-granularity time series. It realizes production characteristic modeling and multi-granularity load modeling by setting up a data preprocessing module, a production characteristic construction module and a multi-granularity load construction module. By setting up a load forecasting module, it realizes multi-task weighted loss, improves the load forecasting accuracy, and ensures the forecasting reliability during critical periods. It is particularly suitable for power management scenarios in industrial parks with complex production scheduling modes and has high engineering application value. Attached Figure Description

[0021] Figure 1 The flowchart illustrates the industrial park load prediction method that integrates production characteristics and multi-granularity time series provided in this embodiment of the invention. Detailed Implementation

[0022] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0023] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] Example 1: This embodiment of the invention discloses a method for predicting the load of industrial parks that integrates production characteristics and multi-granularity time series data. (Refer to...) Figure 1 As shown, the specific steps include the following:

[0025] S1, perform data preprocessing on the historical load data of the industrial park for the target time period, as well as the meteorological data and production day data for the target time period, to obtain the processed historical load data and the meteorological data and production day data for the target time period.

[0026] S2, construct a production feature set for the target time period based on the processed meteorological data and production day data for the target time period;

[0027] S3, construct a multi-granularity load sequence for the historical time period based on the processed historical time period load data;

[0028] S4. Based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period, generate the predicted load sequence of the target time period using the trained industrial park load prediction model.

[0029] Specifically, in step S1, this embodiment collects three core data sources—load data, meteorological data, and daily production data—through standardized communication protocols and interfaces. The collection frequency, data format, and transmission method of each type of data are adapted according to the actual system configuration of the park.

[0030] Since the intraday sampling intervals and collection timestamps of the three types of data sources may differ, time alignment is required before performing data validation and feature construction. The specific method for time alignment is as follows:

[0031] For time alignment of load data for historical periods and meteorological data for target periods, the meteorological data is mapped to intraday time points of the load data using the timestamp of the load data as a reference. The choice of interpolation method is determined according to the characteristics of the meteorological parameters. For continuously changing parameters such as temperature and humidity, linear interpolation or spline interpolation is used, while for cumulative parameters such as precipitation, cumulative allocation method is used.

[0032] For time alignment of load data for historical time periods with production day data for target time periods, discrete production status information is extended into a continuous sequence. For shift information, the sampling time point within each load day is mapped to the corresponding shift identifier according to the shift schedule. For production status identifiers, the forward filling method is used to extend the status value to the next status change point.

[0033] After time alignment is completed, the three types of data form a unified intraday sampling time index.

[0034] To eliminate interference from sensor malfunctions, communication interruptions, manual input errors, and abnormal operating conditions, this implementation method performs a three-level verification mechanism on the load data of historical time periods: physical constraint verification, production time load correlation verification, and meteorological load correlation verification. Among them, the production time load correlation verification and meteorological load correlation verification use production time data and meteorological data of historical time periods.

[0035] Physical constraint verification is achieved using the following formula:

[0036] ,

[0037] in, express Load sampling value at time 10:00 express The rated capacity of the transformer at any given time. express The physical constraint verification results at each moment. express The load sample value at any given time is verified by physical constraints. express The load sample value at time 1 failed the physical constraint verification. This indicates the preset safety factor;

[0038] Production time-load correlation verification is achieved through the following formula:

[0039] ,

[0040] in, express Production day type at any time Label indicating a date of production stoppage or maintenance. express Load sampling value at time 10:00 This represents the historical average load for the same period. Represents the abnormal threshold coefficient. This indicates the results of the production time-load correlation verification. express The load sample value at any given time is verified by correlation with the production time load. , express The load sample value at any given time failed the production time load correlation verification.

[0041] Meteorological load correlation verification is achieved through the following formula:

[0042] ,

[0043] in, express Temperature at any time Indicates the high temperature threshold. This represents the average load during historical high-temperature periods. Represents the abnormal threshold coefficient. This indicates the results of the meteorological load correlation verification. express The load sample values ​​at any given time are verified through meteorological load correlation. , express The load sample value at that time did not pass the meteorological load correlation verification.

[0044] To address data loss caused by communication interruptions, equipment failures, system maintenance, or other reasons during data acquisition, a differentiated completion strategy is adopted. The design principle of the completion strategy is to utilize the periodicity and spatial correlation of the data to ensure that the completed data is as close as possible to the true value.

[0045] Missing data completion for historical load data is achieved using the following formula:

[0046] ,

[0047] in, Indicates the moment of load loss. Indicates time The completed load sample value, This indicates the preset first historical reference period. This indicates the preset second historical reference period. This indicates the preset third historical reference period. Indicates time The load sampling value, Indicates time The load sampling value, Indicates time The load sampling value, This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient. .

[0048] Missing data completion for meteorological data that coincides with the intraday load sampling time is achieved using the following formula:

[0049] ,

[0050] in, Indicates the time when weather is missing. Indicates time The completed meteorological values, Indicates the number of adjacent weather stations. Indicates the sequence number of adjacent meteorological stations. For the first Adjacent weather stations Weather values ​​at any given time This represents the historical meteorological values ​​for the same period at the current weather station. This is a preset flag.

[0051] This embodiment employs a sliding window combined with statistical outlier detection methods to identify and replace load anomaly noise data. Outlier detection and replacement for historical time periods are achieved through the following formula:

[0052] ,

[0053] ,

[0054] ,

[0055] ,

[0056] in, Indicates the half width of the sliding window. Indicates the area within the sliding window relative to The time offset at any given moment. express Load sampling value at time 10:00 express The mean of the sliding window load sampling points at time t. express The standard deviation of the sliding window load sampling points at time t. Indicates the standard deviation multiple threshold. Indicates outlier markers, express The load sampling point at time t is an outlier. express The load sampling point at time t is not an outlier. express The load sample value after replacement at time , express The interpolated load sample value is replaced at a specific time.

[0057] In step S2, production characteristics are the core feature set that distinguishes the method in this embodiment from general load forecasting methods. They directly reflect the driving effect of industrial park production activities on load. The construction process of the production feature set is as follows:

[0058] S2.1, using one-hot encoding, the day type of the production day data for the processed target time period is divided into multiple categories. Each category corresponds to a binary flag bit. Day type categories include weekdays, weekends, statutory holidays, adjusted workdays, and shutdown / maintenance days. For each time point, the day type is determined based on the date information, the flag bit of the corresponding category is set to one, and the remaining flag bits are set to zero, thus obtaining the time feature set. Time feature set The expression is as follows:

[0059] ,

[0060] in, Indicates the day, express The day type encoding vector of the day. Indicates a workday flag. This indicates that the day is a workday. This indicates it is not a working day. Indicates the weekend marker. It means it's the weekend. This indicates it's not the weekend. Indicates the location of statutory holidays. This indicates that it is a statutory holiday. This indicates that it is not a statutory holiday. This indicates the marker for adjusted workdays. This indicates it is a day off in lieu of leave. This indicates that it is not a day off in lieu of leave. This indicates the date of production stoppage or maintenance. This indicates a day of production shutdown and maintenance. This indicates that it is not a day for production shutdown and maintenance.

[0061] S2.2 Meteorological features are used to capture the impact of environmental factors on the load. The original meteorological features, including temperature, relative humidity, and light intensity, are standardized to eliminate dimensional effects. In this embodiment, the processed meteorological data for the target time period are standardized using Z-scores to obtain the meteorological feature set. :

[0062] ,

[0063] in, Indicates time meteorological feature vectors, Indicates temperature. Indicates relative humidity. Indicates the light intensity value;

[0064] S2.3 uses one-hot encoding to classify the intraday shift types of the production day data for the target time period into multiple categories, with each category corresponding to a binary flag bit. Taking a three-shift system as an example, three flag bits are set for the morning shift, afternoon shift, and night shift to obtain the shift feature set. :

[0065] ,

[0066] in, express The time-based train number encoding vector. Indicates the location of the early morning shift. It indicates that it is the early shift. This indicates it's not the early shift. This indicates the location of the middle class marker. It indicates that it is the middle class. This indicates it's not the middle class. Indicates the location of the night shift marker. It indicates that it is the night shift. This indicates it's not the night shift;

[0067] S2.4 Calculate the cumulative daily production time based on the processed production day data for the target time period. Daily production saturation :

[0068] ,

[0069] ,

[0070] in, express The cumulative production time per day express The daily production schedule Indicates the production period. Indicates production period The cumulative production time, express Daily production saturation Indicates the maximum daily production time;

[0071] S2.5, integrate the above-mentioned production characteristics into a production feature vector to obtain the production feature set for the target time period. The expression is as follows:

[0072] ,

[0073] in, express day The production feature vector at time step.

[0074] In step S3, to balance high-frequency fluctuations and long-term trends, this embodiment constructs input sequences with multiple time granularities. The design principle of the multi-granularity sequences is that different granularities capture load patterns at different time scales, and the dominant granularity is adaptively selected through a fusion mechanism.

[0075] S3.1, a fine-grained sequence is constructed using the original sampled data. The sequence length is determined based on the forecast requirements and computing resources. The fine-grained sequence retains high-frequency fluctuation information of the load and is used to capture rapid changes at the minute or hour level, such as load mutations caused by the start-up and shutdown of large equipment or shift switching.

[0076] S3.2, perform time aggregation on the original data to construct a medium-granularity sequence. The aggregation method is to take the average of several consecutive sampling points. The size of the aggregation window is determined according to the original sampling interval. The medium-granularity sequence smooths out random noise and is used to capture hourly load change patterns, such as load fluctuation patterns within a shift.

[0077] S3.3, perform daily aggregation on the raw data to construct a coarse-grained sequence. The aggregation method is to take the average or peak value of all sampling points each day. The coarse-grained sequence is used to capture the daily load change pattern, such as the difference between weekdays and weekends, monthly adjustments to production plans, etc.

[0078] S3.4 The three granularity sequences have different temporal resolutions and sequence lengths, requiring alignment to support fusion. The alignment method involves mapping sequences of different granularities to a unified time index, and using interpolation or forward padding to fill in missing time points, resulting in multi-granularity load sequences for historical time periods with consistent time indices. ;in, , Indicates fine granularity. Indicates medium particle size. Indicates coarse-grainedness.

[0079] In step S4, the industrial park load forecasting model includes a multi-scale feature extraction and fusion module, a time-series dependency modeling module, an attention mechanism module, and an output module. Based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period, the trained industrial park load forecasting model generates a predicted load sequence for the target time period, including:

[0080] S4.1, the multi-granularity load sequence of the historical time period is input into the multi-scale feature extraction and fusion module: multi-scale features of each granularity are extracted using parallel temporal convolutional layers to obtain multi-scale feature map sequences of each granularity; scale fusion is performed on the multi-scale feature map sequences of each granularity using a feature fusion layer to obtain fused feature map sequences of each granularity; the fused feature map sequences of each granularity are obtained by the following formula:

[0081] ,

[0082] in, This represents the sequence of fused feature maps at each granularity. , Indicates fine granularity. Indicates medium particle size. Indicates coarse grain size. This indicates a feature fusion operation. This indicates a channel-level concatenation operation. Indicates the first scale. Indicates the second scale. Indicates the third scale. This represents the sequence of first-scale feature maps for each granularity. The sequence of second-scale feature maps representing each granularity. This represents the sequence of third-scale feature maps for each granularity.

[0083] S4.2, the fused feature map sequences of each granularity are input into the temporal dependency modeling module for long-term bidirectional temporal dependency extraction to obtain the hidden state sequences of each granularity; the temporal dependency modeling module of this embodiment captures the long-term bidirectional temporal dependencies of each granularity sequence based on the bidirectional long short-term memory network Bi-LSTM. The Bi-LSTM structure consists of a forward LSTM layer and a backward LSTM layer. The forward LSTM processes the input sequence in chronological order, and the backward LSTM processes it in reverse order. The two share parameters or are trained independently.

[0084] For each granularity fusion feature sequence, it is input into Bi-LSTM. The hidden state of the forward LSTM at the current time step is calculated from the current input, the hidden state at the previous time step, and the cell state. The hidden state of the backward LSTM at the current time step is calculated based on the reverse time sequence. The bidirectional hidden state sequence is concatenated to obtain the final output.

[0085] S4.3, input the hidden state sequences and production feature sets of each granularity into the attention mechanism module: based on the hidden state sequences of each granularity, calculate the attention weight of each granularity at each time step using the first-layer attention mechanism, and perform a weighted summation of the hidden state sequences of each granularity based on the attention weights at each time step to obtain the context vector of each granularity in the target time period:

[0086] ,

[0087] ,

[0088] ,

[0089] in, Represents the context vector at each granularity of the target time period. Indicates the target time period. express Context vectors at each granularity at any given time. Indicates the time step of each particle size The hidden state, Indicates the length of the time step. Indicates the time step of each particle size Attention weights This represents the softmax activation function. This represents the learnable parameter matrix of the first-layer attention mechanism. Represents the hidden state dimension;

[0090] The context vectors at each granularity of the target time period are concatenated with the production feature set of the target time period. The concatenated vector is then input into the second-layer attention mechanism to obtain the comprehensive context vector of the target time period.

[0091] ,

[0092] ,

[0093] in, The comprehensive context vector representing the target time period. Context vectors representing each granularity of the target time period The result of concatenating the production feature set with the target time period. express The first in One characteristic, express Attention weights This represents the learnable parameter matrix of the second-layer attention mechanism. express The feature dimensions.

[0094] S4.4, using the output module to map the comprehensive context vector of the target time period to the predicted load sequence of the target time period:

[0095] ,

[0096] in, This represents the predicted load sequence for the target time period. This represents the weight matrix of the output module. This indicates the bias term of the output module.

[0097] To improve the model's generalization ability and prevent overfitting, this implementation adopts a multi-task learning framework, dividing the total loss into the main task loss, the first auxiliary task loss, and the second auxiliary task loss; the loss function of the industrial park load forecasting model is as follows:

[0098] ,

[0099] in, Indicates the total loss. Indicates loss of the main task. Indicates the main task loss weight. This indicates the loss of the first auxiliary task. This indicates the loss weight of the first auxiliary task. This indicates the loss of the second auxiliary mission. Indicates the loss weight of the second auxiliary task. , , ;

[0100] The primary task is to predict the load sequence at several future time points. The loss function for the primary task is the time-weighted mean square error, with weights configured as previously described. The primary task's weight coefficient accounts for the highest proportion of the total loss, ensuring its priority optimization. Its loss function applies different weights to the error at each time point. The primary task loss is obtained using the following formula:

[0101] ,

[0102] ,

[0103] in, Indicates the target time period. Indicates the main task is Weighting coefficients at time points, This represents the weighting coefficient for peak production periods. This represents the weighting coefficient for the normal runtime segment. This represents the weighting coefficient for the trough period. , express Forecast load value at time of day express The actual load value at any given time;

[0104] The first auxiliary task is to predict the total electricity consumption for a future period. This task provides global constraints to ensure the accuracy of the total predicted amount. The loss function for the first auxiliary task is the square of the total electricity consumption prediction error, and the difference between the predicted total and the actual total is calculated. The loss of the first auxiliary task is obtained through the following formula:

[0105] ,

[0106] in, This represents the cumulative predicted load for the target time period. This represents the cumulative actual load over the target time period.

[0107] The second auxiliary task is to predict the peak load and its occurrence time in the future. This task provides extreme value constraints to ensure the accuracy of the peak prediction. The loss function of the second auxiliary task includes two terms: peak error and peak time error, which are calculated as the difference between the predicted peak and the actual peak, and the difference between the predicted peak time and the actual peak time, respectively. The loss of the second auxiliary task is obtained by the following formula:

[0108] ,

[0109] in, This indicates the peak predicted load for the target time period. This represents the peak actual load during the target time period. express The moment of appearance express The moment of appearance This represents the error scaling factor.

[0110] This embodiment employs an adaptive learning rate optimizer, such as the Adam optimizer. The initial learning rate is configured based on model complexity and data size. The learning rate decay strategy uses step decay or cosine annealing to gradually reduce the learning rate during training to converge to a better solution. The batch size is configured based on GPU memory capacity and training stability; an excessively large batch size leads to insufficient GPU memory and decreased generalization ability, while an excessively small batch size leads to training instability and slow convergence. The maximum number of training epochs is determined based on the validation set performance. An early stopping mechanism is enabled, stopping training when the validation set loss does not decrease for several consecutive epochs to prevent overfitting. Dropout and L2 regularization are used to prevent overfitting. The Dropout ratio is configured based on model complexity, and the L2 regularization coefficient is configured based on weight decay requirements. Historical data is divided into training, validation, and test sets. The training set is used for model parameter optimization, the validation set for hyperparameter tuning and early stopping judgment, and the test set for final performance evaluation.

[0111] In this embodiment, the predicted load sequence for the target time period is further corrected for meteorological and historical errors, and then subjected to physical constraint limiting to obtain the final load prediction result. The specific process is as follows:

[0112] S5.1 Obtain the weather forecast data for the day to be predicted, calculate its weather similarity with historical days, select several historical days with the highest similarity, and use the model prediction error patterns of these historical days to correct the initial prediction.

[0113] S5.1.1, the similarity of meteorological characteristics between the predicted day and historical days is calculated using a distance metric. The distance metric can be Euclidean distance, Manhattan distance, or Mahalanobis distance. Different weighting coefficients can be set for different meteorological parameters to reflect the relative importance of each parameter to the load. Before distance calculation, the meteorological parameters need to be standardized to eliminate the influence of dimensions. The mathematical expression of meteorological characteristic similarity is as follows:

[0114] ,

[0115] in, express The Middle A historical day, Representing historical days Normalized weights, Indicates the date to be predicted. With History Day The similarity of their weather patterns Indicates the date to be predicted With History Day The similarity of their weather patterns The dimension of meteorological features Indicating meteorological characteristics Custom weights for dimensions Indicates the date to be predicted The Dimensional meteorological characteristic values, Representing historical days The Dimensional meteorological characteristic values;

[0116] S5.1.2 Sort the historical days in ascending order of meteorological distance, and select a number of historical days with the smallest distance to form a set of similar days. The number of similar days is determined based on the scale of historical data and computing resources. For the set of similar days, it is necessary to ensure that the day type is consistent with the day to be predicted, such as matching weekdays with weekdays and matching weekends with weekends.

[0117] S5.1.3 converts distance into weights, with smaller distances resulting in larger weights. The conversion method uses an exponential decay function. The distance is scaled by a popularity coefficient and then exponentially calculated. The popularity coefficient controls the degree of influence of distance on the weights; the larger the popularity coefficient, the more significant the influence of distance. The weights are normalized and then summed to one. The mathematical expression for similarity weight calculation is as follows:

[0118] ,

[0119] in, The first one that is closest in weather similarity to the day to be predicted. A collection of historical days, express The Middle A historical day, Representing historical days Normalized weights;

[0120] S5.1.4, for each similar historical day, calculate the model's prediction error sequence for that day. The error sequence is the difference between the actual load sequence and the model's original prediction sequence. Multiply the error sequences of each similar historical day by their corresponding weights and sum them to obtain the meteorological correction. Add the meteorological correction to the initial prediction sequence to obtain the meteorologically corrected prediction sequence. The mathematical expression for calculating the meteorological correction is as follows:

[0121] ,

[0122] in, express Forecast load value at time of day express Weather correction for each moment;

[0123] S5.2, retrieve several historical days from the historical data that most closely resemble the shape of the load curve the day before the forecast date, and use the model prediction error patterns of these historical days for secondary correction; the mathematical expression for the historical error correction amount is as follows:

[0124] ,

[0125] in, Indicates the attenuation coefficient. This indicates the load most similar to the load of the day before the forecast date. The similarity of loads across a set of historical days is measured using Dynamic Time Warped Distance (DTW). As a preset value, express The Middle A historical day The actual load value at any given time. express The Middle A historical day Forecast load values ​​at any given time;

[0126] S5.3, based on meteorological corrections and historical error corrections, the following corrected forecast load values ​​are obtained:

[0127] ,

[0128] in, express The predicted load value after time correction.

[0129] S5.4, the following physical constraints are applied to the corrected predicted load values:

[0130] ,

[0131] in, express The final predicted load value at time , This indicates taking the minimum value. This represents the preset safety factor. This indicates the rated capacity of the transformer.

[0132] In summary, the industrial park load forecasting method proposed in this embodiment integrates production characteristics and multi-granularity time series. It introduces unique features of industrial parks, such as daily type coding, shift characteristics, cumulative production duration, and production saturation, to effectively capture the impact of production plans on load. It constructs three-level time series of fine, medium, and coarse granularity, and through attention mechanism fusion, it can simultaneously capture minute-level fluctuations, hourly patterns, and daytime trends. Differentiated weights are set for peak production periods, regular operation periods, and off-peak periods to prioritize the forecast accuracy of critical periods. Combined with meteorological correction, historical error correction, and physical constraint verification, it ensures that the forecast results are accurate and reasonable.

[0133] Example 2: Based on the same inventive concept as Example 1, this embodiment of the invention discloses an industrial park load forecasting system that integrates production characteristics and multi-granularity time series, comprising:

[0134] The data preprocessing module is used to preprocess the load data of the industrial park for historical time periods, as well as the meteorological data and production day data for the target time period, to obtain the processed load data of the historical time periods, as well as the meteorological data and production day data for the target time period.

[0135] The production feature construction module is used to: construct a production feature set for the target time period based on the processed meteorological data and daily production data for the target time period;

[0136] A multi-granularity load sequence construction module is used to: construct a multi-granularity load sequence for a historical time period based on the processed historical time period load data;

[0137] The load forecasting module is used to generate a predicted load sequence for the target time period based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period using a trained industrial park load forecasting model.

[0138] In summary, the industrial park load forecasting system proposed in this embodiment, which integrates production characteristics and multi-granularity time series, achieves production characteristic modeling and multi-granularity load modeling by setting up a data preprocessing module, a production characteristic construction module, and a multi-granularity load construction module. By setting up a load forecasting module to implement multi-task weighted loss, the system improves the accuracy of load forecasting and ensures the reliability of forecasting during critical periods. It is particularly suitable for power management scenarios in industrial parks with complex production scheduling modes and has high engineering application value.

[0139] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0140] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0141] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0142] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0143] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0144] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting the load of industrial parks that integrates production characteristics and multi-granularity time series, characterized in that, include: Data preprocessing is performed on the historical load data of the industrial park, as well as the meteorological data and production day data for the target time period, to obtain the processed historical load data, as well as the meteorological data and production day data for the target time period. Based on the processed meteorological data and production day data for the target time period, a production feature set for the target time period is constructed. Construct a multi-granularity load sequence for the historical time period based on the processed historical time period load data; Based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period, the predicted load sequence of the target time period is generated using the trained industrial park load prediction model.

2. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, Data preprocessing is performed on historical load data for the industrial park over a specific period, as well as meteorological and daily production data for the target period, including: Perform physical constraint verification, missing data completion, and outlier detection and replacement on load data for historical time periods; Using the intraday load sampling timestamp as the target time, the meteorological data for the target time period is interpolated and filled, and the production day data for the target time period is forward-filled to obtain meteorological data and production day data consistent with the intraday load sampling time. Missing data were filled in for meteorological data that coincided with the intraday load sampling time; Physical constraint verification of load data for historical time periods is achieved using the following formula: ， in, express Load sampling value at time 10:00 express The rated capacity of the transformer at any given time. express The physical constraint verification results at each moment. express The load sample value at any given time is verified by physical constraints. express The load sample value at time 1 failed the physical constraint verification. This indicates the preset safety factor; Missing data completion for historical load data is achieved using the following formula: ， in, Indicates the moment of load loss. Indicates time The completed load sample value, This indicates the preset first historical reference period. This indicates the preset second historical reference period. This indicates the preset third historical reference period. Indicates time The load sampling value, Indicates time The load sampling value, Indicates time The load sampling value, This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient. ; Outlier detection and replacement of load data for historical time periods is achieved using the following formula: ， ， ， ， in, Indicates the half width of the sliding window. Indicates the area within the sliding window relative to The time offset at any given moment. express Load sampling value at time 10:00 express The mean of the sliding window load sampling points at time t. express The standard deviation of the sliding window load sampling points at time t. Indicates the standard deviation multiple threshold. Indicates outlier markers, express The load sampling point at time t is an outlier. express The load sampling point at time t is not an outlier. express Load sampling value after replacement at time, express The interpolated load sample value is replaced at a specific time. Missing data completion for meteorological data that coincides with the intraday load sampling time is achieved using the following formula: ， in, Indicates the time when weather is missing. Indicates time The completed meteorological values, Indicates the number of adjacent weather stations. Indicates the sequence number of adjacent weather stations. For the first Adjacent weather stations Weather values ​​at any given time This represents the historical meteorological values ​​for the same period at the current weather station. This is a preset flag.

3. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 2, characterized in that, It also includes verifying the correlation between production time and load and the correlation between meteorological load and load data for historical time periods; The production time-load correlation verification is achieved through the following formula: ， in, express Production day type at any time Label indicating a date of production stoppage or maintenance. express Load sampling value at time 10:00 This represents the historical average load for the same period. Represents the abnormal threshold coefficient. This indicates the results of the production time-load correlation verification. express The load sample value at any given time is verified by correlation with the production time load. , express The load sample value at any given time failed the production time load correlation verification. The meteorological load correlation verification is achieved through the following formula: ， in, express Temperature at any time Indicates the high temperature threshold. This represents the average load during historical high-temperature periods. Represents the abnormal threshold coefficient. This indicates the results of the meteorological load correlation verification. express The load sample values ​​at any given time are verified through meteorological load correlation. , express The load sample value at that time did not pass the meteorological load correlation verification.

4. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, The step of constructing a production feature set for the target time period based on the processed meteorological data and production day data for the target time period includes: One-hot encoding is used to divide the day type of the production day data for the target time period into multiple categories, with each category corresponding to a binary flag bit, thus obtaining a time feature set. ;in, , Indicates the day, express The day type encoding vector of the day. Indicates a workday flag. This indicates that the day is a workday. This indicates it is not a working day. Indicates the weekend marker. It means it's the weekend. This indicates it's not the weekend. Indicates the location of statutory holidays. This indicates that it is a statutory holiday. This indicates that it is not a statutory holiday. This indicates the marker for adjusted workdays. This indicates a day off in lieu of leave. This indicates that it is not a day off in lieu of leave. This indicates the date of production stoppage or maintenance. This indicates a day of production shutdown and maintenance. This indicates that it is not a day for production shutdown and maintenance. The processed meteorological data for the target time period are standardized using Z-scores to obtain the meteorological feature set. ;in, , Indicates time meteorological feature vectors, Indicates temperature. Indicates relative humidity. Indicates the light intensity value; One-hot encoding is used to classify the intraday shift types of the production day data for the target time period into multiple categories, with each category corresponding to a binary flag bit, thus obtaining the shift feature set. ;in, , express The time-based train number encoding vector. Indicates the location of the early morning shift. It indicates that it is the early shift. This indicates it's not the early shift. This indicates the location of the middle class marker. It indicates that it is the middle class. This indicates it's not the middle class. Indicates the location of the night shift marker. It indicates that it is the night shift. This indicates it's not the night shift; Calculate the cumulative daily production time based on the processed production day data for the target time period. Daily production saturation ; Based on time feature set Meteorological feature set Class Feature Set Daily cumulative production time Daily production saturation Constructing a set of production characteristics ; The cumulative daily production time Calculated using the following formula: ， in, express The cumulative production time per day express The daily production schedule Indicates the production period. Indicates production period The cumulative production time; The daily production saturation Calculated using the following formula: ， in, express Daily production saturation Indicates the maximum daily production time; The production feature set of the target time period The expression is as follows: ， in, express day The production feature vector at time step.

5. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, The step of constructing a multi-granularity load sequence for a historical time period based on the processed historical time period load data includes: According to the preset fine-grained aggregation window size, the mean value of the load sampled values ​​in each window of the processed historical time period load data is calculated using a sliding window method to obtain the initial fine-grained load sequence; According to the preset medium-granularity aggregation window size, the mean value of the load sampled values ​​in each window of the processed load data is calculated using a sliding window method to obtain the initial medium-granularity load sequence; According to the preset coarse-grained aggregation window size, the mean value of the load sampled values ​​in each window of the processed load data is calculated using a sliding window method to obtain the initial coarse-grained load sequence; Using the time index of the initial fine-grained load sequence as the target time series, interpolation / forward padding is performed on the initial medium-grained load sequence and the initial coarse-grained load sequence to obtain a multi-grained load sequence for historical time periods with consistent time indices. ;in, , Indicates fine granularity. Indicates medium particle size. Indicates coarse-grainedness.

6. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, The industrial park load forecasting model includes a multi-scale feature extraction and fusion module, a time-series dependency modeling module, an attention mechanism module, and an output module. The step of generating a predicted load sequence for the target time period using a trained industrial park load prediction model, based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period, includes: The multi-granularity load sequence of the historical time period is input into the multi-scale feature extraction and fusion module: the multi-scale features of each granularity are extracted using a parallel temporal convolutional layer to obtain a multi-scale feature map sequence of each granularity; the multi-scale feature map sequence of each granularity is fused using a feature fusion layer to obtain a fused feature map sequence of each granularity. The fused feature map sequences of each granularity are input into the temporal dependency modeling module for long-term bidirectional temporal dependency extraction to obtain the hidden state sequences of each granularity. The hidden state sequences and production feature sets of each granularity are input into the attention mechanism module: based on the hidden state sequences of each granularity, the attention weight of each granularity at each time step is calculated using the first-layer attention mechanism, and the hidden state sequences of each granularity are weighted and summed according to the attention weights at each time step to obtain the context vector of each granularity in the target time period; the context vector of each granularity in the target time period is concatenated with the production feature set of the target time period and then input into the second-layer attention mechanism to obtain the comprehensive context vector of the target time period. The output module maps the integrated context vector of the target time period to the predicted load sequence of the target time period.

7. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 6, characterized in that, The fused feature map sequences at each granularity are obtained using the following formula: ， in, This represents the sequence of fused feature maps at each granularity. , Indicates fine granularity. Indicates medium particle size. Indicates coarse grain size. This indicates a feature fusion operation. This indicates a channel-level concatenation operation. Indicates the first scale. Indicates the second scale. Indicates the third scale. This represents the sequence of first-scale feature maps for each granularity. The sequence of second-scale feature maps representing each granularity. The sequence of third-scale feature maps representing each granularity; The context vectors for each granularity of the target time period are obtained using the following formula: ， ， ， in, Represents the context vector at each granularity of the target time period. Indicates the target time period. express Context vectors at each granularity at any given time. Indicates the time step of each particle size The hidden state, Indicates the length of the time step. Indicates the time step of each particle size Attention weights This represents the softmax activation function. This represents the learnable parameter matrix of the first-layer attention mechanism. Represents the hidden state dimension; The comprehensive context vector for the target time period is obtained using the following formula: ， ， in, The comprehensive context vector representing the target time period. Context vectors representing each granularity of the target time period The result of concatenating the production feature set with the target time period. express The first in One characteristic, express Attention weights This represents the learnable parameter matrix of the second-layer attention mechanism. express Feature dimensions; The predicted load sequence for the target time period is obtained using the following formula: ， in, This represents the predicted load sequence for the target time period. This represents the weight matrix of the output module. This indicates the bias term of the output module.

8. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, The loss function of the industrial park load forecasting model is as follows: ， in, Indicates the total loss. Indicates loss of the main task. Indicates the main task loss weight. This indicates the loss of the first auxiliary task. This indicates the loss weight of the first auxiliary task. This indicates the loss of the second auxiliary mission. This indicates the loss weight of the second auxiliary task. , , ; The main task loss is obtained using the following formula: ， ， in, Indicates the target time period. Indicates the main task is Weighting coefficients at time points, This represents the weighting coefficient for peak production periods. This represents the weighting coefficient for the normal runtime segment. This represents the weighting coefficient for the trough period. , express Forecast load value at time of day express The actual load value at any given time; The loss of the first auxiliary task is obtained by the following formula: ， in, This represents the cumulative predicted load for the target time period. This represents the cumulative actual load over the target time period. The loss of the second auxiliary task is obtained by the following formula: ， in, This indicates the predicted peak load for the target time period. This represents the peak actual load during the target time period. express The moment of appearance express The moment of appearance This represents the error scaling factor.

9. The industrial park load forecasting method integrating production characteristics and multi-granularity time series as described in claim 1, characterized in that, The predicted load sequence for the target time period is corrected for meteorological and historical errors, and then subjected to physical constraint limiting to obtain the final load prediction result. The physical constraint limit is achieved through the following formula: ， in, express The final predicted load value at time , This indicates taking the minimum value. express The predicted load value after time correction. This represents the preset safety factor. Indicates the rated capacity of the transformer; The corrected predicted load value It can be obtained through the following formula: ， in, express Forecast load value at time of day express Weather correction for each moment. express Historical error correction amount at any given time; The meteorological correction It can be obtained through the following formula: ， ， ， in, The first one that is closest in weather similarity to the day to be predicted. A collection of historical days, express The Middle A historical day, Representing historical days Normalized weights, Representing historical days exist The actual load value at any given time. Representing historical days exist Forecast load value at time of day This represents the natural exponential function. Indicates the heat coefficient, Indicates the date to be predicted With History Day The similarity of their weather patterns Indicates the date to be predicted With History Day The similarity of their weather patterns The dimension of meteorological features Indicating meteorological characteristics Custom weights for dimensions Indicates the date to be predicted The Dimensional meteorological characteristic values, Representing historical days The Dimensional meteorological characteristic values; The historical error correction amount It can be obtained through the following formula: ， in, Indicates the attenuation coefficient. This indicates the load most similar to the load of the day before the forecast date. The similarity of loads across a set of historical days is measured using Dynamic Time Warped Distance (DTW). As a preset value, express The Middle A historical day The actual load value at any given time. express The Middle A historical day Forecast load value at any given time.

10. An industrial park load forecasting system integrating production characteristics and multi-granularity time series, characterized in that, include: The data preprocessing module is used to preprocess the load data of the industrial park for historical time periods, as well as the meteorological data and production day data for the target time period, to obtain the processed load data of the historical time periods, as well as the meteorological data and production day data for the target time period. The production feature construction module is used to: construct a production feature set for the target time period based on the processed meteorological data and daily production data for the target time period; A multi-granularity load sequence construction module is used to: construct a multi-granularity load sequence for a historical time period based on the processed historical time period load data; The load forecasting module is used to generate a predicted load sequence for the target time period based on the multi-granularity load sequence of the historical time period and the production feature set of the target time period using a trained industrial park load forecasting model.

Images