An Architecture to Improve Energy-Related Time-Series Model Validity Based on the Novel rMAPE Performance Metric

In this paper, an architecture based on computational intelligence for time series modeling is proposed to guarantee the automatic adjustability of trained models no matter the dynamic behavior of the modeled phenomena. Time series are widely used to plan and execute operational and strategic tasks related to the need of forecasting phenomena. Several conventional and non-conventional techniques have been studied for time series modeling. However, the model performance and metrics are affected by non-stationary behaviors. In addition, determining effectively when a model fails can be problematic because the Mean Absolute Percentage Error (MAPE) metric does not necessarily reveal changes in the model predicted curve. Therefore, a novel metric to assess the performance is proposed; and then, an effective maintenance routine for the time-series model is properly devised. Thus, an auditor is created to identify when a model must be updated before losing forecast performance. Hence, using the defined rMAPE performance metric, the auditor output trustworthy detects if the updating process does not achieve better performance, and if replacing a time-series model is required. It is important to note that the devised scheme counts with several assemblies in a local knowledge base. The intelligent system allows building time-series models automatically considering exogenous variables such as weather, calendar, and statistical transformations that can lead to the number of models required for a particular application. The proposed approach has been experimentally tested for power consumption and energy price via simulation. The forecasting results showed an improvement in the MAPE of up to 23% in the tests performed.