A Hybrid Model for Multi-step Time Series Prediction Based on Granular Computing and LSTM

Abstract

Multi-step time series forecasting based on iterative strategies often suffers from severe error accumulation, which significantly degrades long-horizon prediction performance. To address this issue, this paper proposes a hybrid forecasting framework that integrates granular computing with Long Short-Term Memory (LSTM) networks. First, the original time series is transformed into a sequence of temporal information granules, each represented by a triplet consisting of cardinality, mean, and standard deviation, which respectively characterize data density, trend level, and fluctuation intensity. Second, three parallel LSTM networks are constructed to independently predict these granular components, enabling effective feature decoupling and reducing mutual interference during the learning process. Finally, a degranulation mechanism is employed to reconstruct the predicted granules into the original time series, allowing multiple future points to be forecasted within a single prediction step and thereby mitigating error propagation. Extensive experiments on diverse public benchmarks, including classical time series datasets and multi-city PM2.5 concentration data, demonstrate that the proposed framework consistently outperforms traditional statistical methods, standard LSTM models, and existing granulation-based approaches, with im-proved accuracy under long-horizon fore-casting scenarios.