CN-122021740-A - Training method of prediction model, prediction method of park electricity consumption and system thereof

CN122021740ACN 122021740 ACN122021740 ACN 122021740ACN-122021740-A

Abstract

The invention provides a training method of a prediction model, a prediction method of park electricity consumption and a system thereof, wherein the training method comprises the steps of constructing an improved TimeGAN model, replacing an original embedder in a TimeGAN model by a first combined network and replacing an original generator in a TimeGAN model by a second combined network by the improved TimeGAN model, enabling the first combined network and the second combined network to have the same structure, comprising a Multi-scale TCN network and a transform network, inputting second-order historical data of the park electricity consumption into the improved TimeGAN model to obtain sample data, and training a time-cycle neural network model through the second-order historical data and the sample data to obtain the prediction model. The method and the device can effectively solve the problem of poor prediction performance caused by the lack of a large amount of historical data.

Inventors

FU SIXIANG
LI ZEXIAN
DUAN XIAOYONG
CHEN ZHILIE

Assignees

深圳市九牛一毛智能物联科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A method of training a predictive model, the method comprising: constructing an improved TimeGAN model, wherein the improved TimeGAN model is obtained by replacing an original embedder in a TimeGAN model by a first combined network and replacing an original generator in the TimeGAN model by a second combined network, the first combined network and the second combined network have the same structure and comprise a Multi-scale TCN network and a transform network, the first combined network is a combined network obtained through first-order historical data training, and the Multi-scale TCN network and the transform network in the second combined network are untrained combined networks; Inputting second-order historical data of the electricity consumption of the park into the improved TimeGAN model to obtain sample data; And training the time-cyclic neural network model through the second-order historical data and the sample data to obtain a prediction model.
2. The training method of claim 1, wherein said first combined network further comprises an integration network, wherein said set of first order historical data comprises a pair of past historical data and future historical data, said future historical data being data that is closer in time sequence than said past historical data; the step of constructing an improved TimeGAN model includes: Inputting the past history data into the Multi-scale TCN network and the transform network respectively, and integrating the output result of the Multi-scale TCN network and the output result of the transform network into a potential representation through an integration module; inputting the potential representation into the time-cycled neural network model to obtain predicted data; And inputting the predicted data and future historical data corresponding to the past historical data into a loss function module to obtain an error value, and transmitting the error value to the first combination network to adjust parameters of the first combination network.
3. The training method of claim 1, wherein the step of constructing an improved TimeGAN model further comprises: replacing the loss function of the KL divergence used in the TimeGAN model with a target loss function L; L = + ; Wherein, the Represented as a discriminator in the improved TimeGAN model For the second order history data Is used as a means for controlling the speed of the vehicle, Represented as a discriminator in the improved TimeGAN model Sample data generated for the improved TimeGAN model Is not limited to the desired one; , A uniformly distributed random number between 0 and 1, Is the second order history data And sample data Random interpolation between; coefficients that are gradient penalty terms; is a discriminator At random interpolation A gradient thereat.
4. A training method as claimed in any one of claims 1 to 3, wherein the time-cycled neural network model is an LSTM model.
5. A method for predicting electricity consumption of a campus, the method comprising predicting electricity consumption of the campus using a time-cycled neural network model obtained by the training method according to any one of claims 1 to 4.
6. A training system for a predictive model, the training system comprising: A building module configured to build an improved TimeGAN model, wherein the improved TimeGAN model is obtained by replacing an original embedder in a TimeGAN model by a first combined network and replacing an original generator in the TimeGAN model by a second combined network, the first combined network and the second combined network are identical in structure and comprise a Multi-scale TCN network and a transducer network, the first combined network is a combined network obtained through first-order historical data training, and the Multi-scale TCN network and the transducer network in the second combined network are untrained combined networks; the input module is configured to input second-order historical data of the power consumption of the park into the improved TimeGAN model to obtain sample data; And the training module is configured to train the time-cycle neural network model through the second-order historical data and the sample data to obtain a prediction model.
7. The training system of claim 6, wherein said first combination network further comprises an integration network, wherein said set of first order historical data comprises a pair of past historical data and future historical data, said future historical data being data that is closer in time sequence than said past historical data; the construction module comprises: an integration unit configured to input the past history data to the Multi-scale TCN network and the transform network, respectively, and integrate an output result of the Multi-scale TCN network and an output result of the transform network into one potential representation through an integration module; a predictive training unit configured to input the potential representation to the time-cycled neural network model resulting in predictive data; And the adjusting unit is configured to input the predicted data and future historical data corresponding to the past historical data into a loss function module to obtain an error value, and transmit the error value to the first combination network to adjust parameters of the first combination network.
8. A system for predicting electricity consumption of a campus, the system comprising a time-cycled neural network model obtained by the training method according to any one of claims 1 to 4.
9. An electronic device, the electronic device comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the training method of any one of claims 1 to 4.
10. A computer readable storage medium storing computer instructions which, when executed by a processor, implement the training method of any one of claims 1 to 4.

Description

Training method of prediction model, prediction method of park electricity consumption and system thereof Technical Field The invention relates to the technical field of data prediction, in particular to a training method of a prediction model, a prediction method of park electricity consumption and a system thereof. Background Along with the promotion of smart city and green garden construction, the accuracy of garden power consumption prediction directly influences energy utilization efficiency and running cost. The current mainstream prediction methods mainly comprise traditional statistical methods, machine learning methods such as support vector machines, random forests and the like, and deep learning methods such as long-term and short-term memory networks and the like. In practical applications, however, campus electricity usage prediction faces significant small sample data bottlenecks. On one hand, the newly built park lacks long-term historical electricity consumption accumulation, and can only acquire short-term data of a few months or even a few weeks, and on the other hand, partial old parks have the problems of failure of data acquisition equipment, damage of a data storage system and the like, so that the historical electricity consumption data is lost or incomplete. In addition, the traditional statistical method and machine learning method are easy to generate the problems of over fitting of the model and poor generalization capability under a small sample scene. Meanwhile, the existing deep learning method can capture time sequence characteristics of data, but relies on a large number of labeling samples to perform parameter optimization, and under small sample data, model training is easy to sink into local optimization, and association relations among the characteristics cannot be accurately mined. Therefore, a technical solution capable of mining the electric quantity characteristics and improving the prediction accuracy under small sample data is needed. Disclosure of Invention In order to solve the problems, the training method of the prediction model, the prediction method of the power consumption of the park and the system thereof can effectively solve the problem of poor prediction performance caused by the lack of a large amount of historical data by improving the TimeGAN model through the Multi-scale TCN network and the Transformer network. In a first aspect, the present invention provides a training method of a prediction model, the training method comprising: The method comprises the steps of constructing an improved TimeGAN model, wherein an improved TimeGAN model is obtained by replacing an original embedder in a TimeGAN model by a first combined network and replacing an original generator in a TimeGAN model by a second combined network, and the first combined network and the second combined network have the same structure and comprise a Multi-scale TCN network and a transform network; inputting second-order historical data of the electricity consumption of the park into an improved TimeGAN model to obtain sample data; and training the time-loop neural network model through the second-order historical data and the sample data to obtain a prediction model. Optionally, the first combining network further comprises an integration network, wherein the set of first-order historical data comprises a pair of past historical data and future historical data, and the future historical data is data which is closer to the current time than the past historical data in time sequence; the step of constructing an improved TimeGAN model includes: The method comprises the steps of respectively inputting past historical data into a Multi-scale TCN network and a transducer network, and integrating an output result of the Multi-scale TCN network and an output result of the transducer network into a potential representation through an integration module; inputting the potential representation into a time-cycled neural network model to obtain predicted data; The prediction data and the future history data corresponding to the past history data are input into a loss function module to obtain an error value, and the error value is transmitted to a first combination network to adjust parameters of the first combination network. Optionally, the step of constructing the improved TimeGAN model further includes: replacing the loss function of the KL divergence used in the TimeGAN model with a target loss function L; L =+; Wherein, the Distinguishing in TimeGAN model represented as improvementFor second order history dataIs used as a means for controlling the speed of the vehicle,Distinguishing in TimeGAN model represented as improvementSample data generated for improved TimeGAN modelIs not limited to the desired one;, A uniformly distributed random number between 0 and 1, Is second order history dataAnd sample dataRandom interpolation between; coefficients that are gradient penalty terms; is a discrimin