Two level neuro-functional forecaster: A novel dynamic hybridization for functional data forecasting

Abstract

With the advancement of technology, time series data are automatically collected without human intervention. As the data collection process becomes effortless, the next change encountered is to identify the best method to forecast time series data with high accuracies. In this regard, hybrid approaches have gained much attention where the strengths of two approaches can be combined to lessen the weaknesses of each individual approach. When exploring the features of time series data, some depict repetitive patterns and also data can be observed at several levels. The repeating curves can be considered as the higher level, whereas each individual observation can be considered as the lower level. Thus, in order to handle the data in a more effective way, the series can be handled at two levels by giving prominence to the features of each level separately. This paper proposes a novel algorithm named two level neuro-functional forecaster, which is capable of handling data at two levels, hybridizing a statistical approach with an artificial intelligence approach, to gain high accuracy levels. In addition, as this approach handles data at two levels, data sparsity at a particular level can be accommodated at the other level. To apply this algorithm to a real world dataset, electricity demand data in Sri Lanka was considered where the series consisted of daily load curves with repetitive pattern across the days. The proposed hybrid algorithm, outperforms the two approaches when individually used, with a MAPE of 3.324% for a year.