Pseudo-correlation problem and its solution for the transfer forecasting of short-term natural gas loads

Abstract

Considering the information protection of users and system failure of natural gas enterprises, complete large-scale data are difficult to obtain, which is a critical issue for data-driven models applied to short-term load forecasting (STLF). Transfer forecasting has been widely used for STLF because it can significantly improve forecasting accuracy when using scarce data. Pseudo-correlation is a latent issue that significantly affects the accuracy of transfer forecasting and can even generate negative transfers. However, to the best of our knowledge, this phenomenon has not been investigated previously. In this paper, we first identify and define pseudo-correlation, and then propose a novel correlation coefficient algorithm to reduce the risk of negative transfer as a result of pseudo-correlation. We have implemented the proposed algorithm into the transfer forecasting of short-term loads within Greece's National Natural Gas Transmission System. Experimental results demonstrate that when negative transfer occurs, the probability of negative transfer being caused by pseudo-correlation is higher than 40%. Additionally, as the amount of missing data in the target domain increases, this probability also increases. The proposed method can effectively reduce the risk of negative transfers and improve the error in forecasting model by as much as 30%. Our insights into pseudo-correlation and its impact on transfer forecasting may provide forecasters of natural gas supply systems with a helpful method for achieving accurate STLF.