THATSN: Temporal hierarchical aggregation tree structure network for long-term time-series forecasting

Fan Zhang, Min Wang, Wenchang Zhang, Hua Wang

Abstract: The performance of time-series forecasting, which is crucial for predicting future values from historical data, improves with an accurate representation of time-series data. An accurate representation aids in adopting effective strategies to address future uncertainties and reduce the risks associated with planning and decision-making. However, most studies in long-term time-series forecasting integrate features across all time steps, complicating the representation of relationships among cyclical patterns in a series. This paper introduces the Temporal Hierarchical Aggregation Tree Structure Network (THATSN), which is a model that focuses on dynamic modeling over time. We transform time-series data into sequences with multiple cyclical patterns and model the interrelationships among these features to generate a hierarchical tree structure. We design a tree-structured long short-term memory network that functions as a gated adaptive aggregator to process the features of learned child nodes. This aggregator, which can train parameters, adaptively selects child node information benefiting the parent node. This method enhances information usage within the tree and captures cyclical information dynamically. Experiments validate the effectiveness of THATSN, demonstrating its capacity to express cyclical relationships in time-series data. The model exhibits state-of-the-art forecasting performance across several datasets, achieving an overall 15% improvement in MSE, thereby establishing its robustness in long-term forecasting.