4.1.8. Visual Assessment of Predicted Values of the Different Methods

Figure 4 presents the target and predicted values generated by the different models. It is immediately obvious that the graphs are practically overlapped, which further confirms that the models achieved similar performance levels. Essentially, there was only very little difference between the predictive values and the target data prior to the 120th h (i.e., day 5), beyond which a significant error difference was observed. This can be explained noting that a stable pattern existed within the first 5 days, followed by some drop in the target demand level between the 6th and 7th day (i.e., from 120 to 168 h), a period which was not properly tracked by the different methods. This implies that the different ML methods may perform best under conditions with well-defined patterns, and otherwise under heavy stochastic conditions.
