Multivariate Time Series Forecasting Based on Fuzzy Information Granulation and LSTM

Effectively mining features from multivariate time series data to accurately realize multivariate time series forecasting has become a challenging topic. In order to better extract the representative information in multivariate time series, this paper proposes a new prediction model based on fuzzy information granulation and long short-term memory network, which realizes the fine-grained prediction of multivariate time series from series to points. The model firstly partitions multivariate time series into a set of time windows. Secondly, in order to capture the expected information hidden in the time series, we use triangular fuzzy granulates to granulate the fuzzy information of each sub-time series of multivariate time series. Furthermore, the maximum value (upper bound of the interval), the average value and the minimum value (lower bound of the interval) of each sub-time series can be extracted after granulation to construct the information grain for multivariate time series prediction. Finally, it is combined with the long short-term memory network to predict the target variable in multivariate time series. The proposed model is validated using energy time series, and the results show that the method is more accurate in predicting the real-time energy demand of the grid.