Multiple Wavelet Convolutional Neural Network for Short-Term Load Forecasting

Zhifang Liao, Haihui Pan, Xiaoping Fan, Yan Zhang, Li Kuang

Abstract Although the accuracy of load forecasting has been studied by many works, the actual deployability of a model is rarely considered. In this work, we consider the actual deployability of a model from four aspects: 1) the prediction performance of the model; 2) the robustness of the model; 3) the dependence of the model on external data; and 4) the storage size of the model. From these four aspects, we propose a multiple wavelet convolutional neural network (MWCNN) for load forecasting. On two public data sets, we verified the performance and robustness of the MWCNN. The MWCNN only uses load data, and the storage size of the model is only 497 kB, which shows that MWCNN has good deployability. In addition, our MWCNN prediction results are interpretable. The experimental results show that the MWCNN can effectively capture the periodic characteristics of load data.

Introduction: The Deployability Challenge

Short-term load forecasting (STLF) is a cornerstone of modern smart grid management, enabling the balance of electricity supply and demand. While deep learning has pushed the boundaries of forecasting accuracy, many state-of-the-art models remain impractical for real-world deployment. This paper addresses the "deployability gap" by focusing on four critical dimensions: prediction performance, robustness to noise, independence from external data sources (like weather), and model storage size.

Traditional models often rely on external variables such as temperature or social calendars, which may not be consistently available in isolated or secure power systems. Furthermore, massive neural networks are difficult to embed into edge devices. To solve these issues, the authors introduce the Multiple Wavelet Convolutional Neural Network (MWCNN).

Methodology: Wavelet-Enhanced CNNs

The MWCNN architecture leverages a two-stage pipeline designed for efficiency and resilience.

  1. Wavelet Reconstruction (WR): Instead of processing raw, noisy time-series data, the model uses Discrete Wavelet Transform (DWT) to decompose signals into various frequency components. By zeroing out high-frequency noise and reconstructing the signal, the model receives a "cleaned" input that emphasizes underlying trends.
  2. Compact CNN Architecture: The core of the model is a 16-layer CNN. To minimize the parameter count, the authors replaced traditional fully connected layers with Global Average Pooling, resulting in a model that requires only 497 KB of storage.
  3. Ensemble Strategy: To ensure stability across different data patterns, the MWCNN employs an ensemble approach. It trains multiple parallel branches using different wavelet families (e.g., Daubechies and Symlets) and averages their forecasts.
The ensemble MWCNN architecture combines multiple wavelet reconstruction branches to produce a stable forecast.

Experimental Results

The MWCNN was rigorously tested on two public benchmarks: the Independent System Operator New England (ISO-NE) and the North American Utility (NAU) datasets.

Performance and Stability

The model demonstrated superior convergence and accuracy compared to standard benchmarks. On the ISO-NE dataset, the MWCNN ensemble achieved a Mean Absolute Percentage Error (MAPE) of 0.32%, significantly outperforming traditional Artificial Neural Networks (ANNs) which hovered around 0.62%.

Comparison of Test MAPE across different models shows the superior accuracy and stability of the MWCNN variants over 1200 epochs.

Robustness and Compactness

A key finding was the model's resilience. When Gaussian noise was added to simulate sensor errors, the MWCNN's error increased by a mere 0.03%. This robustness, combined with its tiny footprint (33,639 parameters), makes it an ideal candidate for deployment on resource-constrained hardware within the grid.

Interpretability: The 24-Hour Pulse

Beyond raw metrics, the authors used "Integral Gradients" to peek inside the black box. The analysis revealed that the MWCNN naturally prioritizes historical data points from exactly 24 hours prior to the target time. This indicates that the model successfully captures the inherent daily periodicity of human electricity consumption without being explicitly programmed to do so.

Conclusion

The MWCNN represents a shift toward pragmatic AI in power systems. By achieving state-of-the-art accuracy using only historical load data and maintaining a minimal computational footprint, it proves that high-performance forecasting does not require massive models or complex external data integrations. Future research may extend this approach to longer forecasting horizons or more volatile microgrid environments.

Summary updated

Abstract

Although the accuracy of load forecasting has been studied by many works, the actual deployability of a model is rarely considered. In this work, we consider the actual deployability of a model from four aspects: 1) the prediction performance of the model; 2) the robustness of the model; 3) the dependence of the model on external data; and 4) the storage size of the model. From these four aspects, we propose a multiple wavelet convolutional neural network (MWCNN) for load forecasting. On two public data sets, we verified the performance and robustness of the MWCNN. The MWCNN only uses load data, and the storage size of the model is only 497 kB, which shows that MWCNN has good deployability. In addition, our MWCNN prediction results are interpretable. The experimental results show that the MWCNN can effectively capture the periodic characteristics of load data.
BibTeX
@article{liao2020multiple,
  author = {Zhifang Liao and Haihui Pan and Xiaoping Fan and Yan Zhang and Li Kuang},
  title = {{Multiple Wavelet Convolutional Neural Network for Short-Term Load Forecasting}},
  journal = {IEEE Internet of Things Journal},
  year = {2020},
  doi = {10.1109/jiot.2020.3026733},
}
Figure