Interpolating Missing Values In Complex Energy Systems Using Deep Learning And Causal Analysis

With the emergence of the big data era,the energy system has become one of its primary fields,garnering considerable attention.This system encompasses energy consumption,production,storage,and other related aspects,all of which generate crucial data required for efficient management and decision-making.Unfortunately,missing values may occur in the data due to factors such as equipment failure,sensor failure,or data acquisition errors.Such data discrepancies can significantly impair the integrity and reliability of the data,resulting in an inaccurate reflection of the energy system’s operation status.This,in turn,can affect the monitoring and scheduling of the energy system.By imputing the missing values,the data’s accuracy and integrity can be improved,leading to better data quality and more precise operation status reflection of the energy system.Ultimately,this can help operators more effectively monitor and dispatch the energy system.LSTM is a type of recurrent neural network that introduces three gate controllers(input gate,forgetting gate,and output gate)to effectively capture long-term dependencies in a sequence.When interpolating missing values,LSTM mainly infers and fills them in by learning and modeling sequence data.Attention mechanism is a method of weighted aggregation of information achieved by assigning different weights to different parts of input data.In missing value interpolation,the attention mechanism can calculate the similarity between the missing value and other data points to determine the information used to fill the missing value.Based on the above issues and solutions,the main objective of this paper is as follows:(1)Based on LSTM,this paper proposes the use of causal analysis to conduct in-depth learning on multivariate data in the energy system and supplement missing values using the obtained results.Firstly,we rebalance the sample,and then a model based on multivariate LSTM is built.Causal analysis is used to optimize the in-depth learning optimizer,remove unexpected influence factors in the learning process,and obtain the direct impact of stable deflection on the optimization direction.This approach weakens the pseudo-correlation Backdoor Shortcut between the eigenvalues and the stable deflection and eliminates the influence of the stable deflection on the eigenvalues by combining it with the placebo effect.Finally,the harmful factors are subtracted from the eigenvalues to obtain the value of removing the harmful factors,and the model is optimized to achieve better optimization results.This method solves the problems of under-fitting the head data and over-fitting the tail data in machine learning,and attempts to cut off the problem of the machine always involuntarily taking shortcuts in machine learning.The experiments are conducted in the multivariable energy system dataset,and the results show that the problem of missing value interpolation converging to the real value has achieved higher accuracy.(2)Starting with the self-attention mechanism,this paper introduces the Seq2 Seq method to analyze the structure of the Transformer model and improve it to establish the code-decode deep learning model FX＿trans suitable for high-precision energy data interpolation.Through comparative experiments with multiple models,the characteristics and applicable scenarios of each model are analyzed respectively.Through ablation experiments on individual models,the influence mode of each component on the model is analyzed to find the interpolation method with high accuracy and robustness suitable for missing data of complex energy systems.The experimental results show that the improvement of the Transformer structure can significantly reduce the error rate and improved computing speed compared to traditional deep learning models.Finally,the effectiveness of FX＿trans is verified through experiments on a real dataset.