Dynamic fusion LSTM-Transformer for prediction in energy harvesting from human motions

Time series forecasting is critical in many fields, especially for predicting electrical signal characteristics in power systems. Anticipating future fluctuations optimizes resource allocation, improves efficiency, ensures stability, and allows timely responses to issues. In the context of energy harvesting from human motion, forecasting time series data enables identifying movement modes and adjusting device parameters to maximize efficiency. While Long Short-Term Memory (LSTM) networks and Transformer models are effective for time series forecasting, they each have limitations: Transformer may miss local features, and LSTMs can overfit. This paper proposes a Dynamic Fusion LSTM-Transformer model, which combines LSTM’s short-term dependency capture with Transformer’s long-range modeling. The model uses a dynamic fusion mechanism to adjust the weights of LSTM and Transformer based on input data. Experimental results across five time series datasets demonstrate that our model outperforms standalone LSTM and Transformer models with improvements of 0.2-1.9% in MAE, 0.1-3.4% in MSE, and 1.0–3.2% in R2 metrics. Statistical significance tests (p<0.05) confirm the effectiveness of our dynamic fusion strategy.This study enhances time series prediction accuracy and provides new insights for improving energy conversion efficiency in wearable human motion devices.