TL;DR
This paper explores how different signal decomposition techniques can improve short-term forecasting of reference evapotranspiration (ET₀) using Long Short-Term Memory (LSTM) networks. The study finds that while some decomposition methods (VMD, EWT) yield the highest accuracy on test data, others (EMD, ESMD) offer better long-term stability and robustness, crucial for real-world applications in arid regions.
What Problem Does This Paper Address?
Accurate prediction of reference evapotranspiration (ET₀) is vital for managing water resources, especially in arid regions where water is scarce and highly variable. ET₀, a key indicator of water demand by crops, is influenced by numerous complex, non-linear, and non-stationary factors. Traditional prediction models struggle with these characteristics. While deep learning models like LSTM have shown promise, they can still miss embedded features in time series data, leading to prediction inaccuracies. This paper aims to address this by investigating how combining various signal decomposition methods with LSTM can enhance ET₀ forecasting accuracy and stability.
Methods
The core of this research lies in evaluating eight different "decomposition-hybrid" models. The process involves:
- Decomposition: Each of the eight chosen decomposition algorithms (EMD, EEMD, CEEMDAN, VMD, LMD, ESMD, DWT, EWT) is applied to the ET₀ time series. These algorithms break down the complex, non-stationary ET₀ signal into several simpler, more stable components (Intrinsic Mode Functions - IMFs and residuals). The idea is that these simpler components are easier for a predictive model to learn from.
- Prediction: A Long Short-Term Memory (LSTM) network, a powerful deep learning model for time series, is then used to predict each of these decomposed components.
- Reconstruction: The individual predictions for each component are then combined to generate the final ET₀ forecast.
The study also verifies the accuracy of the decomposition algorithms themselves using an artificial sequence with known components before applying them to the real ET₀ data.
Datasets
The study utilizes a 40-year daily meteorological dataset (1990-2019) from the Yulin meteorological station in Shaanxi Province, China. This dataset includes daily maximum and minimum temperature, sunshine hours, relative humidity, and wind speed at 10m height. The ET₀ values were calculated using the FAO-56 Penman-Monteith equation. The dataset was split into training (80%), testing (20%), and an out-of-sample validation (last 10 days) set to simulate real-world forecasting scenarios. No new benchmark datasets were introduced.
Key Insights and Ideas
- Decomposition is Key, but the "Right" Decomposition Matters: Simply applying any decomposition method isn't enough. The choice of decomposition technique significantly impacts the final prediction's accuracy and, more importantly, its stability.
- High-Frequency Components are a Double-Edged Sword: Methods like VMD and EWT excel at capturing detailed high-frequency information, leading to excellent performance on test data. However, this abundance of high-frequency detail can introduce cumulative errors and lead to overestimation in actual predictions, reducing robustness.
- Stability Over Raw Accuracy for Prediction: For practical forecasting, especially in dynamic environments like arid regions, predictive stability (i.e., consistent accuracy over time and across different prediction horizons) is paramount. EMD and ESMD, while not always achieving the absolute highest R² on test sets, demonstrated superior stability in out-of-sample predictions.
- The "Completeness" of Decomposition is Crucial: The study emphasizes the importance of ensuring that decomposition algorithms accurately reconstruct the original signal without losing information. This was verified using artificial sequences.
Key Results
- Top Performers in Testing: VMD-LSTM (M4) and EWT-LSTM (M8) achieved the highest R² values (0.983 and 0.992, respectively) on the testing set for 5-day ahead forecasting.
- Most Stable Performers in Prediction: EMD-LSTM (M1) and ESMD-LSTM (M6) showed superior predictive stability, with no significant difference between their predictions and actual ET₀ values (p > 0.05) in the out-of-sample validation.
- Accuracy Degradation: All models showed a decrease in accuracy with longer prediction horizons (5, 7, and 10 days). EMD-LSTM (M1) and ESMD-LSTM (M6) exhibited the slowest R² degradation.
- Overall Improvement: Hybrid decomposition models generally improved accuracy by 14.1% to 27.6% compared to the standalone LSTM model.
Limitations and Future Work
- Single Station Study: The study was conducted using data from only one meteorological station. The authors acknowledge that this limits the generalizability of the findings and suggest that future research should include more regions, stations, and meteorological elements to validate the results.
- High-Frequency Component Handling: The paper highlights the challenge of handling excessive high-frequency components generated by some decomposition methods. Future work could focus on bi-decomposition techniques or optimization/screening processes for these high-frequency variables to further improve prediction accuracy and robustness.
- Sequence Length Threshold: Determining the optimal sequence length for high-precision prediction remains an area for future investigation.
- Robustness: Developing more robust prediction models that maintain accuracy across varying prediction times is also a key direction for future research.
Summary updated