Energy-consuming enterprise time series data prediction method and system based on reversible computation

Abstract

The application belongs to the technical field of information processing, and provides a kind of energy-consuming enterprise time series data prediction method and system based on reversible calculation, the energy-consuming enterprise time series data prediction model in the application is a trained deep learning model, the input sample during training includes local time series information and hierarchical time series information;At the same time, the actual time series data is compared with the predicted time series data, and the predicted result is revised according to the comparison result, on the basis of considering the particularity and complexity of hierarchical time series information, through the reversible calculation of the comparison of actual time series data and predicted sequence data, in the high energy-consuming enterprise time series data prediction task, the modeling of long-term dependence relationship, flexibility, adaptation to different time scales, support variable length sequence, multi-task learning and processing of abnormal value and fluctuation can be embodied Better characteristics, selecting root mean square error as loss function can make the prediction model get better constraint on enterprise time series data prediction task.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information processing, and particularly relates to a time series data prediction method and system for energy-consuming enterprises based on reversible computing. BACKGROUND

[0002] In the production process of energy-consuming enterprises, the generated time series data usually presents complex time series data, including time reports, daily reports, weekly reports, monthly reports and annual reports. These data must be accurately and timely collected, integrated and analyzed to support decision-making, resource optimization and industry regulation. However, the processing of time series data involves dynamic updating and correction of data, and in addition to this, it is very difficult to ensure data quality simply by relying on manual data statistics, because the energy consumption data of the enterprise may not be reported in time due to some uncontrollable factors.

[0003] The inventors found that because the data generated by high energy-consuming enterprises in the production process is more complex than general time series data, there are many problems in predicting it. For example, long-term dependence modeling is insufficient, traditional time series prediction methods may have deficiencies in modeling long-term dependence, especially when the energy consumption of high energy-consuming enterprises is affected by multiple factors in the long term, which may lead to difficulties in accurately predicting future consumption trends; limited adaptability to variable length sequences, some traditional methods may have limited adaptability to variable length time series data, while the energy consumption pattern of high energy-consuming enterprises may be affected by factors such as seasonality and production changes, resulting in uncertainty in sequence length; lack of flexibility to adapt to multiple time scales, some traditional methods may lack flexibility and cannot effectively adapt to information on different time scales. SUMMARY

[0004] To solve the above problems, the present application provides a time series data prediction method and system for energy-consuming enterprises based on reversible computing, which helps to improve the prediction performance of the model and better meet the accurate prediction needs of high energy-consuming enterprises for time series data such as energy consumption.

[0005] To achieve the above purpose, the present application is realized by the following technical scheme:

[0006] In a first aspect, the present application provides a time series data prediction method for energy-consuming enterprises based on reversible computing, comprising:

[0007] obtaining enterprise energy consumption data;

[0008] According to the enterprise energy consumption data, and a preset energy consumption enterprise time series data prediction model, predicted time series data is obtained; wherein the energy consumption enterprise time series data prediction model is a trained deep learning model, and the input sample during training includes local time series information and hierarchical time series information; the root mean square error is selected as a loss function;

[0009] The actual time series data is compared with the predicted time series data, and the prediction result is revised according to the comparison result; wherein during the comparison, the continuation is carried out from a smaller granularity to a larger granularity.

[0010] Further, the enterprise energy consumption data includes energy consumption, production output and equipment state.

[0011] Further, the obtained enterprise energy consumption data is preprocessed by deleting variables, filling, identifying abnormal values and normalizing.

[0012] Further, if the missing rate of the variable is greater than a first preset percentage, the coverage is lower than a preset value, and the variable is directly deleted;

[0013] If the missing rate is less than a second preset percentage, the data is filled according to the data distribution; when the data conforms to uniform distribution, the missing data is filled with the mean value; when the data exists in the skewed distribution, the missing data is filled with the median value;

[0014] The outlier detection algorithm is used to monitor and identify abnormal data, and the mean value of the data set is used to replace the identified abnormal values.

[0015] Further, the energy consumption enterprise time series data prediction model uses a trained Informer model;

[0016] The input local time series information includes historical energy consumption data information and other time series data, and the calculation is in time steps, and the calculation method is:

[0017]

[0018]

[0019] Wherein, 2j represents the dimension of even number, 2j+1 represents the dimension of odd number; pos represents the position of data in time series data; d model is the feature dimension after input representation; L x represents the length of input data;

[0020] The hierarchical time series information is represented by using a learnable dot embedding, and the calculation formula of input representation is:

[0021]

[0022] Wherein, p represents p types of global timestamps; and a is a balance scalar mapping and a factor of the size between local / global embedding. is the energy consumption data Context mapping to d model dimensional vector.

[0023] Further, the self-attention mechanism in the Informer model adopts a screening method to screen out Q that meets the uniform distribution with the score of each K;

[0024] Remove data redundancy, give higher weight to dominant attention advantage features, and generate feature mapping in the next layer.

[0025] Further, the loss function calculation formula is as follows:

[0026]

[0027] Wherein, Y i represents the sample true value; represents the sample corresponding prediction value; and n represents the sample quantity.

[0028] In a second aspect, the present application further provides a time series data prediction system for energy-consuming enterprises based on reversible calculation, comprising:

[0029] A data acquisition module configured to acquire enterprise energy consumption data;

[0030] A prediction module configured to obtain predicted time series data according to the enterprise energy consumption data and a preset time series data prediction model for energy-consuming enterprises; wherein the time series data prediction model for energy-consuming enterprises is a trained deep learning model, the input sample during training includes local time series information and hierarchical time series information; and the root mean square error is selected as the loss function;

[0031] A revision module configured to compare the actual time series data with the predicted time series data, and revise the prediction result according to the comparison result; wherein the comparison is continued from a smaller granularity to a larger granularity level.

[0032] In a third aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the time series data prediction method for energy-consuming enterprises based on reversible calculation according to the first aspect.

[0033] In a fourth aspect, the present application further provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the time series data prediction method for energy-consuming enterprises based on reversible calculation according to the first aspect when executing the program.

[0034] Compared with the prior art, the application has the beneficial effects that:

[0035] The time series data prediction model of the high-energy-consuming enterprise in the application is a trained deep learning model, the input sample during training includes local time series information and hierarchical time series information; the root mean square error is selected as the loss function, which is superior to the traditional model, at the same time, the actual time series data is compared with the predicted time series data in a way of continuing from a smaller granularity to a larger granularity, and the predicted result is revised according to the comparison result, on the basis of considering the particularity and complexity of the hierarchical time series information, through the reversible calculation of the comparison of the actual time series data and the predicted time series data, in the time series data prediction task of the high-energy-consuming enterprise, the modeling of long-term dependence, flexibility to adapt to different time scales, support for variable-length sequence, multi-task learning and processing of abnormal values and fluctuations can be embodied, and the selection of the root mean square error as the loss function can make the prediction model better constrained in the enterprise time series data prediction task; the application is helpful to improve the prediction performance of the model and better meet the accurate prediction demand of the high-energy-consuming enterprise for the time series data such as energy consumption. BRIEF DESCRIPTION OF DRAWINGS

[0036] The drawings constituting a part of this embodiment are used to provide further understanding of this embodiment, the illustrative embodiments of this embodiment and the description thereof are used to explain this embodiment, and do not constitute improper limitation on this embodiment.

[0037] Figure 1 The method flowchart of the embodiment 1 of the application;

[0038] Figure 2 The model diagram of the embodiment 1 of the application for predicting the production energy consumption data of the key energy-consuming enterprise based on the Informer model;

[0039] Figure 3 The process flowchart of the time series data prediction system of the key energy-consuming enterprise based on the reversible calculation of the embodiment 1 of the application. DETAILED DESCRIPTION

[0040] The application will be further described below in combination with the drawings and embodiments.

[0041] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0042] Embodiment 1:

[0043] In the information age, data has become the core resource driving innovation and technological development. However, in this vast data lake, the issues of data quality, management, and traceability are increasingly prominent. In the energy sector, in particular, a large number of data sources are pouring in, and how to efficiently collect, store, manage, and analyze these data becomes crucial.

[0044] The time series data of key energy-consuming enterprises usually includes multiple dimensions, covering energy consumption, production output, equipment status, and other aspects. The analysis and management of these multi-dimensional data require different data processing methods and tools. At the same time, the data formats and standards of different enterprises are also different, so efficient management of energy consumption and other time series data of energy-consuming enterprises has become a difficult and key problem.

[0045] The time series data generated in the production process of key energy-consuming enterprises usually presents as complex time series data, including daily, weekly, monthly, and annual reports. These data must be accurately and timely collected, integrated, and analyzed to support decision-making, resource optimization, and industry regulation.

[0046] However, the processing of time series data involves dynamic updating and correction of data, and because the energy consumption data reported by enterprises may not be timely due to some uncontrollable factors, relying solely on manual data statistics makes it difficult to ensure data quality. Therefore, a reversible calculation method is needed to introduce a data traceability and review mechanism to ensure that the source of the data is traceable, the processing of the data is monitorable, and the analysis results of the data are reviewable, which will improve the credibility of the data and the scientificity of the decision-making.

[0047] Because the data generated by high-energy-consuming enterprises in the production process is more complex than general time series data, there are many problems in predicting it. For example, long-term dependence modeling is insufficient, traditional time series prediction methods may have deficiencies in modeling long-term dependence, especially when the energy consumption of high-energy-consuming enterprises is influenced by multiple factors over a long period. This may make it difficult to accurately predict future consumption trends; limited adaptability to variable length sequences, some traditional methods may have limited adaptability to variable length time series data, while the energy consumption patterns of high-energy-consuming enterprises may be affected by factors such as seasonality and production changes, resulting in uncertainty in sequence length; lack of flexibility to adapt to multiple time scales, some traditional methods may lack flexibility and cannot effectively adapt to information on different time scales. Enterprise energy consumption may exhibit different patterns on different time scales such as hours, days, months, and years.

[0048] Based on this, the embodiment provides a time series data prediction method for energy-consuming enterprises based on reversible calculation. The loss function of the Informer model is modified, and the modified Informer model performs better than general traditional models. The method comprises the following steps:

[0049] Obtaining enterprise energy consumption data;

[0050] Obtaining predicted time series data according to the enterprise energy consumption data and a preset time series data prediction model for energy-consuming enterprises, wherein the time series data prediction model for energy-consuming enterprises is a trained deep learning model, the input sample during training includes local time series information and hierarchical time series information, and the root mean square error is selected as the loss function.

[0051] Comparing the actual time series data with the predicted time series data, and revising the prediction result according to the comparison result; wherein, during the comparison, the comparison is continued from a smaller granularity to a larger granularity.

[0052] The method in the embodiment can reflect the modeling of long-term dependence, flexibility in adapting to different time scales, support for variable-length sequences, multi-task learning, and processing of abnormal values and fluctuations, and the like, in the time series data prediction task for high-energy-consuming enterprises, by reversible calculation of the comparison between the actual time series data and the predicted time series data, based on the particularity and complexity of hierarchical time series information. Selecting the root mean square error as the loss function can better constrain the prediction model in the enterprise time series data prediction task. The present application helps to improve the prediction performance of the model and better meet the accurate prediction needs of high-energy-consuming enterprises for time series data such as energy consumption.

[0053] Through the reversible calculation in the method of the embodiment, when the enterprise data is not reported in time, the delayed data needs to be reasonably predicted based on the past data of the enterprise to provide real-time data support for decision-making, resource optimization and industry supervision. However, when the delayed data is reported, the predicted data needs to be revised and recalculated to achieve reversible calculation of the data, and the predicted data is covered and revised using the new data.

[0054] The specific steps of the method of the embodiment are as follows:

[0055] S1, obtaining enterprise energy consumption data:

[0056] Optionally, in the embodiment, time series data generated in the production process of key energy-consuming enterprises is obtained. These data can include information in multiple dimensions, such as energy consumption, production output and equipment status, etc. The following key factors need to be considered for data acquisition:

[0057] Data source identification: Determine which data sources to collect data from, which may include internal production systems, device sensors, production reports, and possibly external data sources.

[0058] Multi-source data integration: In addition to energy consumption data, other time-series data generated during the production process, such as production output and device status, can be integrated. Multi-source data integration can provide more comprehensive information and help improve the accuracy of predictions.

[0059] Real-time data stream processing: Capture and process data streams from various data sources in real time to ensure real-time data.

[0060] S2, data preprocessing step:

[0061] To ensure the credibility of the quality of the collected data and the accuracy of the model prediction, data quality control mechanisms need to be implemented.

[0062] During the process of collecting and reporting data in enterprise statistics, there may be various reasons for data loss and gaps. The main method for handling these missing values is based on the distribution characteristics of the variables and the importance of the variables (information content and prediction ability) using different methods. In this embodiment, the processing of missing values is mainly divided into the following several methods:

[0063] S2.1, delete variable, if the missing rate of the variable is higher, such as more than 80%, the coverage is low, and the importance is low, the variable can be directly deleted.

[0064] S2.2, statistical filling, if the missing rate is less than 95% and the importance is low, then fill according to the data distribution. For data that conforms to uniform distribution, fill the missing values with the mean of the variable. For data with skewed distribution, use the median to fill.

[0065] S2.3, use the Z-score method in the outlier detection algorithm to monitor and identify abnormal data, and use the mean of the data set to replace the identified outliers. Evaluate and monitor the processed data to ensure the quality and accuracy of the processed data. Regularly monitor the performance of the outlier detection algorithm to adapt to changes and drifts in data.

[0066] The calculation formula of Z-score is:

[0067]

[0068] Where x is the sample data; is the mean of the sample data; s is the standard deviation of the sample data.

[0069] S2.4, Considering that different evaluation indicators often have different dimensions and units, which will affect the results of data analysis, in order to eliminate the influence of dimensions between indicators, data normalization processing is needed, and in this embodiment, Min-Max normalization is adopted, and the calculation formula is as follows:

[0070]

[0071] Where, x is the sample data; x min is the minimum value of the sample data; x max is the maximum value of the sample data.

[0072] After the above steps, the data noise needs to be processed. Noise is the random error and variance of the variable, which is the error between the observation point and the true point. In this embodiment, the processing method selected is to perform a binning operation on the data, and then use the average, median or boundary value (select different processing methods according to the sample data distribution) of each bin to replace all the numbers in the bin, which plays a role in smoothing the data.

[0073] These works will ensure that the time series data of key energy-consuming enterprises can be effectively obtained, and provide the necessary data basis for subsequent reversible calculation.

[0074] S3, optionally, input the time series data obtained after preprocessing into the Informer model, output the energy consumption data of the enterprise in a certain period of time in the future, and the work flow is as shown in Figure 2 .

[0075] The Informer model is a typical encoder-decoder structure, in which the encoder receives a representation vector matrix of the data and outputs an encoding information matrix. The encoder is usually composed of multiple Encoder Blocks. The decoder receives the encoding information matrix output by the encoder and restores it as much as possible to the semantics that the original data wants to express. In other words, the encoder is used for feature extraction of the original signal. The decoder is only a reverse component in the training process of the encoder, which helps the encoder to train.

[0076] Input and output formal representation:

[0077] Input:

[0078] Output:

[0079] Where, represents the sample data at time step t; is the output value of the model corresponding to the time step; in the background of the embodiment, it is the prediction value of the enterprise time series data.

[0080] To input the enterprise energy consumption data into the Informer model, the embodiment first needs to construct a proper input representation. For the energy consumption prediction problem, the following information is usually needed:

[0081] Local temporal information: This includes historical energy consumption data information and other temporal data, usually in time steps. This can be represented as a time series, where each time step corresponds to the energy consumption value in that time step. The calculation method is to save the context information by fixed position embedding, and the calculation method is represented by the following formula:

[0082]

[0083]

[0084] wherein, 2j represents the even dimension, 2j+1 represents the odd dimension; pos represents the position of the data in the time series data; d model is the feature dimension after input representation; L x represents the length of the input data.

[0085] Hierarchical temporal information: The Informer model can process information of multiple time scales, such as weeks, months, and years. Therefore, in the embodiment, different scale time series information can be input into the model together to better capture the trends of different time scales. SE (pos) represents the hierarchical temporal information.

[0086] Therefore, the calculation formula of the input representation is as follows:

[0087]

[0088] wherein, p represents p types of global time stamps; α is a balance scalar that balances the size between the local / global embedding and the local / global embedding; is the energy consumption data mapped to a d model dimensional vector using a one-dimensional convolution filter.

[0089] Optionally, the self-attention mechanism in Informer is ProbSparse Self-attention.

[0090] First, the traditional self-attention (self-attention) mechanism input form is (query, key, value), and then the scaled dot product (scaleddot-product) is performed, that is:

[0091]

[0092] wherein, d is the input dimension; the probability form of the attention coefficient of the i-th Query is:

[0093]

[0094] However, considering that when the data set is large, if the conventional Self-Attention is used for calculation, the calculation complexity will reach L x 2 The self-attention mechanism in the Informer model adopts a screening method, so that the calculation complexity is reduced to L x ×logL x The screening mechanism is introduced as follows.

[0095] In the conventional Self-Attention, the inner product of Q and K is the similarity score of Q and K, but there will be some Q, which has a uniform distribution with the score of each K, such Q has little effect on considering global information, so it is necessary to screen out such Q and make appropriate processing on the Q.

[0096] Before screening Q, in order to further reduce the calculation complexity, K will be randomly sampled for screening Q, and then all Q and the sampled K are calculated, as follows.

[0097]

[0098] When is small, it means that Q i has little effect, and Q i will not participate in the calculation of the similarity score. For the data at position i, the mean value is used to reconstruct it.

[0099] Removing data redundancy: In the Informer model, the Self-Attention Distilling operation is proposed. Because there are redundant combinations of V values in the feature map of the encoder, the Distilling operation is used to give higher weights to the dominant features with dominant attention, and generate focused self-attention feature map in the next layer.

[0100] It can be understood that in the embodiment, the distilling operation is used to give higher weights to the dominant features with dominant attention, and generate focus self-attention feature mapping in the next layer. The process of distilling calculation from j to j+1 layer is as follows:

[0101]

[0102] where [·] aB The basic operation in multi-head sparse self-attention and attention block is included; Conv1d(·) represents one-dimensional convolution operation on time series, and ELU is used as the activation function, and finally the maximum pooling operation is performed.

[0103] In order to enhance the robustness of the distilling operation, the Informer model establishes a half copy of the main stack, and gradually reduces the number of self-attention distillation layers, and each time it drops a layer to ensure that the output dimension is aligned. Therefore, the output of connecting all the stacks can get the final hidden representation of the encoder.

[0104] Decoder: Produce long sequence output through a forward process. The Informer model uses a standard decoder structure. The decoder is composed of two identical multi-head attention layers stacked together. The vector representation input into the decoder is calculated as follows

[0105]

[0106] where is the start token, is the placeholder of the target sequence.

[0107] By setting "masked dot-products" to -∞, the masked multi-head attention mechanism is applied to ProbSparse self-attention, and the probability self-attention calculation is performed to prevent each position from paying attention to the next position, avoiding autoregression.

[0108] A fully connected layer finally obtains the final output, and the output dimension is determined by whether the prediction is a single variable prediction or a multi-variable prediction. The method uses a generative structure in the encoder structure, which can generate all prediction sequences at a time, greatly shortening the prediction decoding time.

[0109] Compared with the Informer model, we modify its loss function. In order to better adapt to the task needs of time series data prediction for high-energy-consuming enterprises, we select the root mean squared error RMSE (Root Mean Squared Error) as the loss function to calculate the error between the output of the decoder and the target sequence for back propagation. The loss function calculation formula is as follows:

[0110]

[0111] where, y i represents the sample true value; represents the prediction value corresponding to the sample; n represents the sample number.

[0112] After the model training is completed, the obtained prediction data is well labeled, such as the date to which the prediction data belongs, whether it is prediction data, etc., so as to revise the data when the real data of the enterprise is obtained.

[0113] S4, reversible calculation:

[0114] In this embodiment, reversible calculation is used to ensure the revision of predicted energy consumption data.

[0115] When new actual energy consumption data is available, the previous energy consumption prediction data needs to be revised. This is a key step to realize reversible calculation. The revision process may involve the following aspects:

[0116] S4.1, compare new data and prediction data: compare the new actual time series data with the previously predicted time series data to identify differences and deviations. This helps to determine which parts need to be revised. It should be noted that in step S3, the special nature and complexity of hierarchical time series information have been taken into account, so when comparing new data and prediction data, we decide to start comparing from the smaller granularity, and continue to the larger granularity level.

[0117] S4.2, revise energy consumption prediction:

[0118] According to the new actual data, as well as the date label and type of prediction data, revise the previous energy consumption prediction data, that is, according to different time stamp, when the actual obtained time series data has a large deviation from the predicted data, the actual data will replace the original prediction data.

[0119] S4.3, data storage:

[0120] Store the time series data of the enterprise in the storage system in layers. These data include historical energy consumption data and revised energy consumption prediction data. Store these data according to time sequence, so as to subsequent access, revision and prediction.

[0121] S4.4, keep revision history:

[0122] When revision is performed, the revision history needs to be kept for future audit and analysis. These records should include the time of revision, data before and after revision, etc.

[0123] S4.5, data version management:

[0124] In order to better realize the reversible calculation, and the safety of data, the data needs to be managed by version. Each revision should generate a new version, and ensure that the previous version can be traced back at any time.

[0125] S4.6, update in time:

[0126] The revised data should be updated into the system in time for subsequent decision-making and resource optimization. At the same time, the updated data should also be fed back to the Informer model to continuously improve the accuracy of the prediction.

[0127] Embodiment 2:

[0128] The embodiment provides a time series data prediction system for energy-consuming enterprises based on reversible computing, comprising:

[0129] The data acquisition module is configured to obtain enterprise energy consumption data.

[0130] The prediction module is configured to obtain predicted time series data according to the enterprise energy consumption data and a preset time series data prediction model for energy-consuming enterprises, wherein the time series data prediction model for energy-consuming enterprises is a trained deep learning model, the input sample during training includes local time series information and hierarchical time series information, and the root mean square error is selected as the loss function.

[0131] The revision module is configured to compare the actual time series data with the predicted time series data, and revise the prediction result according to the comparison result, wherein the comparison is performed from a smaller granularity to a larger granularity.

[0132] The working method of the system is the same as the time series data prediction method for energy-consuming enterprises based on reversible computing in Embodiment 1, and will not be repeated here.

[0133] Embodiment 3:

[0134] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps of the time series data prediction method for energy-consuming enterprises based on reversible computing in Embodiment 1.

[0135] Embodiment 4:

[0136] The embodiment provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor realizes the steps of the time series data prediction method for energy-consuming enterprises based on reversible computing in Embodiment 1 when executing the program.

[0137] The above only describes the preferred embodiments of the present embodiment and is not intended to limit the present embodiment. Those skilled in the art can make various modifications and changes to the present embodiment. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present embodiment shall be included in the protection scope of the present embodiment.

Claims

1. A method for predicting time-series data of energy-consuming enterprises based on reversible computation, characterized in that, include: Obtain enterprise energy consumption data; The enterprise energy consumption data includes energy consumption, production output and equipment status, which meets the forecasting needs of high-energy-consuming enterprises for time-series energy consumption data. Based on enterprise energy consumption data and a pre-set time-series data prediction model for energy-consuming enterprises, predicted time-series data is obtained; wherein, the time-series data prediction model for energy-consuming enterprises is a trained deep learning model, and the input samples during training include local time-series information and hierarchical time-series information; the root mean square error is selected as the loss function. The actual time-series data is compared with the predicted time-series data, and the prediction results are revised based on the comparison results. Considering the special nature and complexity of hierarchical time-series information, the comparison is carried out from smaller granularity to larger granularity. Through reversible calculation, when enterprise data is not reported in a timely manner, reasonable predictions are made for the delayed data based on the enterprise's past data. When the delayed data is reported, the predicted data needs to be revised and recalculated to achieve reversible calculation of the data, using new data to overwrite and revise the predicted data. Reversible computation is employed to ensure revisions to predicted energy consumption data. Previous energy consumption predictions need to be revised when new actual energy consumption data becomes available. Specifically: Comparing new and predicted data: New actual time-series data is compared with previously predicted time-series data to identify differences and deviations; comparisons begin at smaller granularities and proceed sequentially to larger granularities; revising energy consumption predictions: Previous energy consumption predictions are revised based on the new actual data, as well as the date tags and types of the predicted data; data storage: The enterprise's time-series data is stored hierarchically in the storage system; maintaining revision history, including the revision time and data information before and after the revision; data version management: each revision should generate a new version, ensuring that previous versions can be reverted to at any time; timely updates: revised data should be updated to the system promptly and also fed back into the Informer model.

2. The method for predicting time-series data of energy-consuming enterprises based on reversible calculation as described in claim 1, characterized in that, The acquired enterprise energy consumption data is preprocessed by deleting variables, filling in data, identifying outliers, and normalizing the data.

3. The method for predicting time-series data of energy-consuming enterprises based on reversible calculation as described in claim 2, characterized in that, If the missing rate of a variable is greater than the first preset percentage, and the coverage is lower than the preset value, the variable will be deleted directly. If the missing rate is less than the second preset percentage, then fill in the missing data according to the data distribution; if the data is uniformly distributed, fill in the missing data with the mean; if the data is skewed, fill in the missing data with the median. An outlier detection algorithm is used to monitor and identify abnormal data, and the mean of the dataset is used to replace the identified outliers.

4. The method for predicting time-series data of energy-consuming enterprises based on reversible computation as described in claim 1, characterized in that, The time-series data prediction model for energy-consuming enterprises uses a pre-trained Informer model. The input local time-series information includes historical energy consumption data and other time-series data, calculated in time steps as follows: in, 2j represents an even-numbered dimension, and 2j+1 represents an odd-numbered dimension; pos represents the position of the data in the time series data; Lx represents the feature dimension after input representation; Lx represents the length of the input data. The hierarchical temporal information is represented using learnable stamp embeddings. The formula for calculating the input representation is: Where p represents p types of global timestamps; A factor to balance the size between scalar mapping and local / global embedding; It uses a one-dimensional convolutional filter to process energy consumption data. Context mapping to A dimensional vector.

5. The method for predicting time-series data of energy-consuming enterprises based on reversible computation as described in claim 4, characterized in that, The self-attention mechanism in the Informer model employs a filtering method to select Q values ​​that conform to a uniform distribution with respect to the scores of each K. Data redundancy is removed, higher weights are assigned to dominant features that attract attention, and feature maps are generated in the next layer.

6. The method for predicting time-series data of energy-consuming enterprises based on reversible computation as described in claim 1, characterized in that, The formula for calculating the loss function is as follows: in, Indicates the true value of the sample; This represents the predicted value corresponding to the sample; n represents the number of samples.

7. A time-series data prediction system for energy-consuming enterprises based on reversible computation, characterized in that, include: The data acquisition module is configured to acquire enterprise energy consumption data. The enterprise energy consumption data includes energy consumption, production output and equipment status, which meets the forecasting needs of high-energy-consuming enterprises for time-series energy consumption data. The prediction module is configured to: obtain predicted time-series data based on enterprise energy consumption data and a preset time-series data prediction model for energy-consuming enterprises; wherein, the time-series data prediction model for energy-consuming enterprises is a trained deep learning model, and the input samples during training include local time-series information and hierarchical time-series information; and the root mean square error is selected as the loss function. The revision module is configured to compare actual time-series data with predicted time-series data and revise the prediction results based on the comparison. Considering the special nature and complexity of hierarchical time-series information, the comparison proceeds from smaller granularities to larger granularities. Through reversible calculation, when enterprise data reporting is untimely, reasonable predictions are made for delayed reporting based on the enterprise's past data. After delayed data is reported, the predicted data needs to be revised and recalculated, achieving reversible calculation of the data, using new data to overwrite and revise the predicted data. Reversible computation is employed to ensure revisions to predicted energy consumption data. Previous energy consumption predictions need to be revised when new actual energy consumption data becomes available. Specifically: Comparing new and predicted data: New actual time-series data is compared with previously predicted time-series data to identify differences and deviations; comparisons begin at smaller granularities and proceed sequentially to larger granularities; revising energy consumption predictions: Previous energy consumption predictions are revised based on the new actual data, as well as the date tags and types of the predicted data; data storage: The enterprise's time-series data is stored hierarchically in the storage system; maintaining revision history, including the revision time and data information before and after the revision; data version management: each revision should generate a new version, ensuring that previous versions can be reverted to at any time; timely updates: revised data should be updated to the system promptly and also fed back into the Informer model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for predicting time-series data of energy-consuming enterprises based on reversible computation as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting time-series data of energy-consuming enterprises based on reversible computation as described in any one of claims 1-6.

Images