[Paper Review] LoadCNN: A Low Training Cost Deep Learning Model for Day-Ahead Individual Residential Load Forecasting

Yunyou Huang, Nana Wang+5 more|arXiv (Cornell University)|Aug 1, 2019

Energy Load and Power Forecasting25 references4 citations

TL;DR

LoadCNN is a lightweight, convolutional neural network model designed for day-ahead individual residential load forecasting with significantly reduced training costs. It achieves state-of-the-art prediction accuracy while reducing training time by 98.1%, energy consumption by 97.8%, and CO₂ emissions by 97.8% compared to existing deep learning models.

ABSTRACT

Accurate day-ahead individual residential load forecasting is of great importance to various applications of smart grid on day-ahead market. Deep learning, as a powerful machine learning technology, has shown great advantages and promising application in load forecasting tasks. However, deep learning is a computationally-hungry method, and requires high costs (e.g., time, energy and CO2 emission) to train a deep learning model, which aggravates the energy crisis and incurs a substantial burden to the environment. As a consequence, the deep learning methods are difficult to be popularized and applied in the real smart grid environment. In this paper, we propose a low training cost model based on convolutional neural network, namely LoadCNN, for next-day load forecasting of individual resident with reduced training cost. The experiments show that the training time of LoadCNN is only approximately 1/54 of the one of other state-of-the-art models, and energy consumption and CO2 emissions are only approximate 1/45 of those of other state-of-the-art models based on the same indicators. Meanwhile, the prediction accuracy of our model is equal to that of current state-of-the-art models, making LoadCNN the first load forecasting model simultaneously achieving high prediction accuracy and low training costs. LoadCNN is an efficient green model that is able to be quickly, cost-effectively and environmentally-friendly deployed in a realistic smart grid environment.

Motivation & Objective

• Address the high computational and environmental costs of training deep learning models for residential load forecasting.
• Develop a model that maintains state-of-the-art prediction accuracy while drastically reducing training time, energy use, and CO₂ emissions.
• Enable practical deployment of deep learning in real-world smart grid applications by minimizing resource consumption.
• Demonstrate that model efficiency and accuracy can be simultaneously achieved without sacrificing forecasting performance.

Proposed method

• Propose a lightweight CNN-based architecture specifically optimized for individual residential load forecasting.
• Use a shallow, efficient network structure to minimize parameters and layers, reducing training complexity.
• Apply convolutional layers to extract temporal patterns from historical load data without requiring deep stacking.
• Optimize hyperparameters and network design through systematic experimentation to balance accuracy and efficiency.
• Train the model on large-scale smart meter datasets to ensure robustness and generalization across diverse residential consumption patterns.
• Benchmark performance against state-of-the-art models (e.g., LSTM, seq2seq, TCN) using standard metrics like RMSE, NRMSE, and MAE.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve state-of-the-art load forecasting accuracy while drastically reducing training cost?
• RQ2 To what extent can model architecture simplification reduce training time, energy consumption, and CO₂ emissions without degrading prediction quality?
• RQ3 Is it possible to maintain high forecasting accuracy using a shallow CNN-based model instead of deeper RNN or transformer-based architectures?
• RQ4 How do training costs (time, energy, emissions) scale across different deep learning models in residential load forecasting?
• RQ5 What is the impact of model depth on forecasting performance and training efficiency in day-ahead load prediction tasks?

Key findings

• LoadCNN reduces training time to approximately 1/54th of state-of-the-art models, achieving 98.1% faster training.
• Energy consumption during training is reduced to about 1/45th of comparable models, representing a 97.8% improvement in energy efficiency.
• CO₂ emissions from training LoadCNN are only about 1/45th of those from other SOTA models, equivalent to reducing emissions by 97.8%.
• Despite its simplicity, LoadCNN achieves prediction accuracy on par with the best-performing models in terms of RMSE, NRMSE, and MAE.
• Deeper models (e.g., 8-layer LSTM, seq2seq) showed worse performance, indicating diminishing returns and overfitting risks with increased depth.
• The study confirms that improving accuracy through architectural depth or hyperparameter tuning alone is ineffective; future gains require richer input features like human activity data.