TY - JOUR
T1 - Aggregation of nonlinearly enhanced experts with application to electricity load forecasting
AU - Incremona, Alessandro
AU - De Nicolao, G.
AU - Fusco, F.
AU - Eck, B. J.
AU - Tirupathi, S.
PY - 2021
Y1 - 2021
N2 - Combining the predictions of different base experts is a well known approach used to improve the accuracy of time series forecasts. Forecast aggregation is becoming crucial in many fields, including electricity forecasting, as Internet of Things and Cloud technology give access to larger numbers of sensor data, time series and predictions from external providers. In this context, it is not uncommon that the failure of some experts causes relevant drops in the performances of the aggregated forecast when classical techniques based on linear averaging are applied. This might be a symptom of suboptimality of the individual experts, that do not fully exploit important predictors, e.g. calendar features that play a major role in the electrical demand profiles. In this work, we therefore present two non-linear strategies to obtain aggregated forecasts, starting from the availability of a set of base experts and the knowledge of some relevant predictor variables. The first approach, called aggregation of enhanced experts (AEE), enhances each individual expert and then feeds the enhanced forecasts into classical linear aggregation techniques. In the second approach, called enhanced aggregation of experts (EAE), the expert forecasts are nonlinearly combined with the predictor variables through an Artificial Neural Network (ANN). The case of missing expert forecasts is also considered via a statistically-based imputation method. The short-term prediction of German electrical load is used as a case study. Twelve base experts are enhanced with respect to calendar features, i.e. daytime and weekday. Compared to state-of-the-art aggregation methods applied to the not-enhanced set of experts, the proposed approaches not only improve the accuracy of aggregated forecast (up to 25% reduction of MAPE and RMSE), but are also robust with respect to missing experts.
AB - Combining the predictions of different base experts is a well known approach used to improve the accuracy of time series forecasts. Forecast aggregation is becoming crucial in many fields, including electricity forecasting, as Internet of Things and Cloud technology give access to larger numbers of sensor data, time series and predictions from external providers. In this context, it is not uncommon that the failure of some experts causes relevant drops in the performances of the aggregated forecast when classical techniques based on linear averaging are applied. This might be a symptom of suboptimality of the individual experts, that do not fully exploit important predictors, e.g. calendar features that play a major role in the electrical demand profiles. In this work, we therefore present two non-linear strategies to obtain aggregated forecasts, starting from the availability of a set of base experts and the knowledge of some relevant predictor variables. The first approach, called aggregation of enhanced experts (AEE), enhances each individual expert and then feeds the enhanced forecasts into classical linear aggregation techniques. In the second approach, called enhanced aggregation of experts (EAE), the expert forecasts are nonlinearly combined with the predictor variables through an Artificial Neural Network (ANN). The case of missing expert forecasts is also considered via a statistically-based imputation method. The short-term prediction of German electrical load is used as a case study. Twelve base experts are enhanced with respect to calendar features, i.e. daytime and weekday. Compared to state-of-the-art aggregation methods applied to the not-enhanced set of experts, the proposed approaches not only improve the accuracy of aggregated forecast (up to 25% reduction of MAPE and RMSE), but are also robust with respect to missing experts.
KW - Enhancement
KW - Forecast aggregation
KW - Neural networks
KW - Missing features
KW - Load demand
KW - Enhancement
KW - Forecast aggregation
KW - Neural networks
KW - Missing features
KW - Load demand
UR - http://hdl.handle.net/10807/295517
U2 - 10.1016/j.asoc.2021.107857
DO - 10.1016/j.asoc.2021.107857
M3 - Article
SN - 1568-4946
VL - 112
SP - N/A-N/A
JO - Applied Soft Computing
JF - Applied Soft Computing
ER -