Abstract
A new variational mode decomposition (VMD) based deep learning approach is proposed in this paper for time series forecasting. Firstly, VMD is adopted to decompose the original time series into several sub-signals. Then, a convolutional neural network (CNN) is applied to learn the reconstruction patterns on the decomposed sub-signals to obtain several reconstructed sub-signals. Finally, a long short term memory (LSTM) network is employed to forecast the time series with the decomposed sub-signals and the reconstructed sub-signals as inputs. The proposed VMD-CNN-LSTM approach is originated from the decomposition-reconstruction-ensemble framework, and innovated by embedding the reconstruction, single forecasting, and ensemble steps in a unified deep learning approach. Empirical results demonstrate that the proposed approach outperforms consistently the benchmark approaches in terms of forecasting accuracy.
Introduction
Time series forecasting—predicting future values based on past data—is notoriously difficult when dealing with "noisy" or non-linear signals like commodity prices or weather patterns. The current state-of-the-art often uses a decomposition-ensemble approach: breaking a complex signal into simpler sub-signals (modes), forecasting each mode individually, and then summing the results.
However, most researchers perform these steps separately. This disjointed approach means the forecasting model cannot influence how the signal is reconstructed, potentially leaving valuable patterns on the table. Furthermore, the "reconstruction" step (deciding which sub-signals belong together) is often handled via manual heuristics rather than learned from the data. This paper proposes a unified framework to bridge this gap.
Methodology
The proposed VMD-CNN-LSTM framework integrates three distinct tools into one end-to-end pipeline:
1. Variational Mode Decomposition (VMD)
VMD is an adaptive signal processing technique that decomposes the original time series into $K$ sub-signals (modes). Unlike older methods, VMD is non-recursive and robust to noise, providing a cleaner input for the neural network layers.
2. Convolutional Neural Network (CNN) for Reconstruction
This is the core innovation of the framework. Instead of manually grouping sub-signals, a CNN layer learns "reconstruction patterns." It treats the decomposed modes as multi-channel inputs and learns the optimal weights to recombine them into feature-rich "reconstructed sub-signals."
3. Long Short-Term Memory (LSTM) for Forecasting
The final stage uses an LSTM network, a recurrent architecture designed for long-term dependencies. The LSTM takes both the raw decomposed modes and the CNN-reconstructed modes as inputs to generate the final forecast.
Because the CNN and LSTM are part of the same computational graph, the model can be trained end-to-end. The reconstruction step is optimized specifically to minimize the final forecasting error.
Experimental Evaluation
The model was evaluated on four real-world datasets representing distinct forecasting challenges:
- Crude Oil Prices (Daily): West Texas Intermediate (WTI) and Brent spot prices (2011–2018), characterized by high economic volatility.
- Wind Speed (15-min intervals): Two datasets from a wind farm in Inner Mongolia, China (2015), representing high-frequency, stochastic physical data.
Key Results and Insights
The proposed VMD-CNN-LSTM consistently outperformed all benchmarks, including standard LSTMs and other VMD-hybrid models (VMD-SVR, VMD-RFR).
- Significant Accuracy Gains: In the WTI Crude Oil dataset, the VMD-CNN-LSTM achieved an out-of-sample RMSE of 0.5162, compared to 1.0466 for a standard LSTM—a reduction in error of approximately 50%.
- The Power of End-to-End Training: Embedding the reconstruction and ensemble steps into a single network allows the model to discover a "multi-dimensional representation" of the original series that standard models cannot capture.
- CNN as a Feature Engineer: The results indicate that the CNN isn't just adding signals; it learns complex spatial-temporal patterns across the different frequency modes generated by VMD. Comparing VMD-CNN-LSTM to VMD-LSTM (which lacks the CNN layer) showed that the CNN-learned features were critical for the final accuracy boost.
Conclusion
By unifying decomposition, reconstruction, and forecasting, the VMD-CNN-LSTM framework provides a robust solution for volatile time series data. While the current study used a relatively simple CNN structure, future work could explore more complex architectures or swap the LSTM for newer models like Transformers to further enhance performance.
Summary updated