New Hybrid Weighted Predictive Data Analysis Approach: Application to Temporal Power Transformer Maintenance Data

This paper proposes a new methodological approach for extracting typical data profiles from time series of power transformer maintenance databases with the aim of preventing future faults. The proposed approach operates in two stages. First, the model formalizes the date variable as a time series and then analyzes the data to identify interdependencies between them. Rio Tinto Alcan's dissolved gas data from transformer T0001 show that H2, CH4, C2H4, C2H6, CO, CO2, and N2 are dependent, while C2H2 and O2 are independent. In the second step, to capture short- and long-term trends, seasonalities, and dependencies in the data on the one hand and to extract non-linear trends and seasonalities for variable forecasting on the other, an ARIMA (Autoregressive Integrated Moving Average) + GES (Generalized Exponential Smoothing) model in combination is applied to the data series. The hybrid ARIMA (2, 1, 1) + GES (0.1, 0.1, 1) model, with a weighting of 0.5, produced errors of 0.27 and 4.5%, respectively, in terms of Mean Absolute Scaled Error (MASE) and Mean Symmetric Absolute Percentage Error (SMAPE). The ARIMA model taken individually gave MASE equals 0.55 and SMAPE equals 8.9%. Similarly, the proposed model is better than the naive model because the MASE is less than 1 (0.27%). The data series was subjected to other forecasting models, and it was found that the model proposed in this article is more accurate given the error results obtained since the smaller the error, the more accurate the forecasting model.