Multi-model Combination Of Short-term Power Load Forecasting Based On LSTM And Informer

Research on energy efficiency in the power sector is important for China to meet the objectives of its dual carbon strategy.Operating and regulating the power system effectively is a key component of energy conservation in the power sector,and a crucial foundation for this is the ability to predict short-term power load demand with accuracy.Operations and maintenance managers can create rational power scheduling plans and maintain the equipment’s smooth operation with the aid of a more precise short-term load forecast.Single-feature input power load forecasting and multi-feature input power load forecasting under various data conditions are two categories for short-term power load forecasting tasks.When only load time series data are available,the former refers to the electric load values as feature input,whereas the latter also includes other data such as temperature,humidity,and electricity price as input features in addition to load data.This thesis conducts research on single-day and multi-day forecasting of short-term electric load from the aspects of electric load forecasting values under different data conditions,which are influenced by single feature factors and multiple feature factors in order to obtain more accurate short-term electric load forecasting results.The following are the primary research findings and contents:(1)The Long Short-Term Memory(LSTM)network is used for load forecasting in single-day power load forecasting with single feature input in order to address the peculiarities of load data temporality and the selection of model network parameters.The LSTM offers distinct model advantages in tackling load forecasting issues due to its robust learning memory and nonlinear processing capability.The IPSO-LSTM short-term electric load forecasting model is created for single-day load forecasting by adjusting the basic PSO algorithm to produce Improved Particle Swarm Optimization(IPSO)and adding adaptive perturbation variables.The model performs well at making predictions,according to experimental findings using real data from campus apartments and open-source data.(2)A single LSTM network paired with parameter optimization makes it harder to perform efficient feature extraction because multiple feature factor input power load forecasting has more intricate relationships between each influencing factor than single feature input load forecasting.This thesis breaks down the original load data into several modal components using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)method to address the issue.Then,the structure-adjusted Convolutional Neural Network(CNN)is used for feature extraction of each component,and subsequently input into the LSTM network for load value prediction to establish the CEEMDAN-CNN-LSTM short-term electric load prediction model.The model performs better in single-day electrical load forecasting with multi-feature factor inputs,according to experimental results using public data with multi-feature factor inputs and data pertaining to actual campus public buildings.(3)When equivalent single-day load forecasting models are produced for two different load forecasting input characteristics separately,a CEEMDAN-Informer-based short-term electric load forecasting model is developed for the actual multi-day load forecasting demand problem.The multiscale feature extraction and sparse self-attentiveness mechanisms of the Informer model make it more effective for longer series of multi-feature input forecasting problems.The findings demonstrate that when the forecast length is 7 days for a full week and 5 days for a weekday period,the CEEMDAN-Informer load forecasting model has a greater forecasting accuracy.In summary,different LSTM and Informer model combinations are chosen and combined with other methods for single-day power load forecasting with a single feature input,single-day power load forecasting with multiple feature inputs,and multi-day power load forecasting with multiple feature inputs,respectively,in this thesis based on the study of the technical basis of shortterm power load forecasting.The predictions on the associated data successfully show how the shortterm load forecasting models developed in this study can be applied to various forecasting tasks.Figure [36] table [18] reference [56]...