CNN-LSTM architectures for non-stationary time series: decomposition approach

Abstract

Time series data, representing sequential observations recorded over time, plays a pivotal role in numerous domains. Extracting valuable insights from time series is crucial for informed decision-making, pattern recognition, and predictive modeling. Traditional forecasting methods, such as moving averages and autoregressive models, have been effective for stationary time series. However, the increasing complexity of time series data across various applications has highlighted the limitations of these methods in capturing seasonal and trend patterns. This paper focuses on the CNN-LSTM model, which combines the strengths of both Convolutional Neural Networks and Long Short-Term Memory (LSTM) networks. The Conv-LSTM architectures leverages CNNs' feature extraction capabilities and LSTM’s ability to capture temporal dependencies. The study aims to evaluate the CNN-LSTM’s different architectures performance in forecasting non-stationary time series. It incorporates a decomposition approach to assess the architecture’s ability to capture trend and seasonal components and utilize them as features for making predictions. The results shed light on the effectiveness of the architectures of CNN combined with LSTM in capturing the different patterns of the data, by exhibiting the lowest MSE and MAE for the trend, seasonal, and residual components.