Genetic algorithm-based cyclic encoding for improved nonlinear regression load forecasting

Load forecasting provides a means to anticipate electricity consumption patterns. The expected patterns, such as load consumption, help stakeholders in power industries attain effective electricity management from the output of the forecasting process. However, optimizing predictor variables of a load forecasting model increases its performance. Scholars agree that, it is more challenging to obtain optimal encoded values when it comes to predictor variable whose values are ordinal. Such ordinal values in short-term load forecasting, for example, include the “type of day” and “time of day” (e.g., weekday, weekend, morning, afternoon, and so on). Several approaches for encoding these values have been proposed, including ordinal, one-hot, binary, and heuristic-based. However, the values encoded using these methods do not significantly improve forecasting accuracy. This paper applies the artificial intelligence technique based on the genetic algorithm (GA) to improve the accuracy of the encoded values for the time of day variable in load forecasting. Paired t-test, Wilcoxon, and Cohen’s d tests are used to evaluate the significance of improving the GA encoding method. Results indicate that the GA encoding approach outperforms by improving the accuracy of the forecasting model by 42.35%, 95.25% (based on MAE), and 37.87%, 94.81% (based on RMSE) over the heuristic and ordinal encoding, respectively. The proposition of the approach to improve load forecasting accuracy in this study provides an optimal numerical representation of cyclic ordinal variables (such as time of the day) directly based on forecasting error minimization.