Missing Data Imputation Based On A Segmented Shape-representation For Energy Data Of Metallurgy Industry

Metallurgy industry is one of the most important industries and has important strategic significance to the country. Because the metallurgical process needs to consume large amounts of energy, the unreasonable production process will not only wastes the energy but also brings serious environment pollution. For the enterprises, the complete and reliable historical data is the foundation for their optimization and scheduling. Since the production process is fairly complex, data collection often results in different kinds of missing phenomenon. At present, the operators usually use their personal experience or by a simpler method such as the average of the data to impute the missing data. However, the experience based method cannot satisfy the accuracy requirement in practice, which will directly lead to the failure of the whole scheduling solution, and bring large adverse effects to the enterprise.Based on the background of the production process of the metallurgy industry, this study has done a lot of researches on their data loss situation and puts forward a missing data imputation way based on a segmented shape-representation method. The method uses historical data to establish a database, adopts the key point and the sliding window technology to reduce the dimension of the data, and represents the target sequence through its trends, amplitude levels and degrees of volatility. At last, this paper uses the above shape representation technology to compare the similarity of target sequence and other sequences in the database and chooses the most similar sequences as the training samples. Then, an echo state network model based on Gaussian process is adopted to implement the filling data.To validate the performance of the proposed method, firstly, the parameter’s influence on the result of sequence segment is analyzed and the segment parameter is determined. Then, this study conducts an experiment on the industrial production data, in which the filled accuracy is compared to some other methods. The experiment result shows that the proposed method exhibits significant improvement to the imputation effect on different types and degrees of missing data.