Hybrid Algorithm for Real-Time Data Forecasting with Density-Based Clustering and Penalty Splines

We propose a hybrid forecasting model that includes time series segmentation, which is based on the DBScan clustering algorithm, and a penalty spline. Each step of the hybrid model required modification for real-time use. The DBScan density-based algorithm was supplemented with technology for working with noise and transition between clusters. A recurrent penalty spline model has low computational complexity and, accordingly, high forecasting speed inside the hybrid model. The forecast efficiency was assessed by a number of indicators that reflect predictive efficiency and extrapolation accuracy of using synthetic, model and real TS. Accuracy indicators (MSPE, MAPE) are below 4%. In comparison with other widely-spread forecasting methods, the proposed hybrid algorithm ranks among the known algorithms.