The time series forecasting research has become very popular in recent years, which affects the economic interests in various industries, for instance, wind power, stock price and electrical load power forecasting. However, the majority of the time series patterns are uncertain, nonlinear and fluctuating that effect on the predicted accuracy of the model. With the evolution of statistical and deep learning models, the problem has been alleviated but it is tedious and lengthy to find the optimal value due to the massive hyperparameter figures in most of the models or methods. Especially, it is time-consuming for the researchers and practitioners who are new in the time series forecasting field. In this research, a novel model is proposed and this model is not only easy to use but also adept in tackling complexly nonlinear systems with large amount of data. Firstly, the similar data is clustered into the group by using the S-means, followed by applying the data fusion method in order to reduce the data dimensionality. Secondly, the whole system can be represented by a small amount of fused data while outputting the prediction result. Finally, the time analysis function is utilized to correct the final prediction result. The proposed model has the following advantages: there are fewer and simpler hyperparameters needed which is more efficient to use during the time adjustive process; the forecasting accuracy is more stable, and the calculation rate is faster due to the non-parametric and non-training. To compare the proposed model with some existing models or methods, the forecasting accuracy of each dataset has a better performance, as it shown in the paper, the forecasting accuracy has improved with 2.5% in the wind speed data; 70% in Google stock price data; 12% in the electricity load data; 31% in the weekly British Pound/US dollar (GBP/USD) exchange rate data, and 24% in the sunspot data.
Read full abstract
Sep 1, 2023
Tim Palmer 
Open Access