A Hybrid Method for Long-Term Demand Forecasting in the Electrical Energy Supply Chain of Basic Metal Production Industries in the Presence of Incomplete Data

The economic growth of any country has a lot to do with the infrastructure of the electrical energy supply chain and the ability to access it at low cost. Increasing the resilience of the electric energy supply chain in order to be able to respond to the real time demand of high-consumption and strategic consumers is a challenge that will not be possible without considering long-term demand forecasting and integrated development planning of this chain. This paper presents a long-term demand forecasting approach in the electrical energy supply chain of Isfahan's Espidan iron stone industries. This approach is a combination of wavelet transform, long short-term memory (LSTM) network and finally integrating the results with data-mining technique based on machine learning. The company studied in this research is one of the main suppliers of raw materials in the supply chain of basic metal production industries and one of the ten energy-intensive industries in the electrical energy supply chain of Isfahan province. The only information available from this company is the daily time series signal of the historical electrical energy demand of this industry in a period of 40 months. The data in the studied time series is interrupted so that only 50% of the data has a value and the remaining 50% is zero. This lack of data and the impossibility of access to supplementary data and effective features for forecasting has reduced the density of data and the possibility of long-term demand forecasting faces more problems than continuous time series. The used statistical analysis showed that the annual and seasonal data do not follow the normal distribution and have high distortion and heterogeneity. The proposed method and its results have been compared with other available approaches. The results of 10 iterations of extreme learning machine methods show that the RELM technique with a high confidence level of 95% is more effective than other machine learning methods and has more accurate results.