Abstract
Accurate electricity consumption forecasting is vital for the stability and efficiency of modern power grids. Traditional point forecasting methods often fail to account for the inherent uncertainties in energy demand, providing only a single estimated value. This paper introduces a hybrid forecasting framework, LASSO-QRNN, which integrates the Least Absolute Shrinkage and Selection Operator (LASSO) with Quantile Regression Neural Networks (QRNN) and Kernel Density Estimation (KDE). By performing intelligent feature selection and modeling the full probability distribution of future consumption, this method significantly improves prediction accuracy and provides a robust measure of uncertainty for decision-makers.
Introduction
Predicting regional electricity consumption is a complex task influenced by a "perfect storm" of socio-economic and environmental factors, such as GDP growth, industrial shifts, population dynamics, and weather patterns. Conventional forecasting models typically face two major hurdles. First is the "curse of dimensionality," where including every possible variable leads to model "clutter," overfitting, and poor out-of-sample performance. Second is the limitation of point predictions; grid operators require more than just a single number—they need to understand the probability of extreme spikes or dips to manage risk effectively. To address these issues, this research proposes a three-stage probabilistic forecasting pipeline that moves beyond single-number estimates.
Methodology: The LASSO-QRNN Framework
The proposed LASSO-QRNN approach consists of three integrated stages designed to refine raw data into actionable probabilistic insights:
- Feature Selection via LASSO: The model utilizes the Least Absolute Shrinkage and Selection Operator to filter high-dimensional datasets. LASSO acts as a smart filter by applying a penalty to the absolute size of regression coefficients, effectively "shrinking" the coefficients of unimportant or redundant variables to zero. This ensures that the neural network focuses only on the most influential drivers, reducing computational noise and preventing the model from learning patterns that do not generalize.
- Probabilistic Modeling with QRNN: Instead of a standard neural network that predicts a conditional mean, the Quantile Regression Neural Network (QRNN) is employed. This network learns the relationship between inputs and specific quantiles (e.g., the 5th, 50th, or 95th percentile) of electricity consumption. By training on multiple quantiles, the model captures non-linear relationships across the entire distribution, allowing it to describe the shape of the uncertainty.
- Continuous Smoothing with KDE: The discrete quantile predictions generated by the QRNN are finally processed using Kernel Density Estimation (KDE). This step transforms the point-wise quantile estimates into a smooth, continuous probability density function (PDF). The resulting curve provides a visual and mathematical representation of future uncertainty, enabling grid operators to assess the likelihood of various consumption levels at a glance.
Experimental Evaluation and Results
The model's performance was validated using historical data from two distinct, real-world regions to ensure its robustness across different economic contexts: Guangdong Province, China (1995–2016) and California, USA (1990–2015).
Key Findings
- Variable Importance: The LASSO stage revealed that the primary drivers of consumption vary significantly by region. In Guangdong, a region of rapid industrialization, the "secondary industry proportion" and "permanent residents" were the most critical factors. In contrast, the mature California market was more heavily influenced by "total energy" and "average electricity price."
- Accuracy Gains: In the Guangdong dataset, the LASSO-QRNN model achieved a Mean Absolute Percentage Error (MAPE) of just 1.21%. The inclusion of LASSO was pivotal; it reduced the Relative Mean Square Error (RMSE) from 1.10% to a staggering 0.05% compared to a standard QRNN without feature selection.
- Robustness in Volatile Markets: California presented a unique challenge due to a notable decline in consumption following the 2008 financial crisis. Despite this "intractable" trend, the model maintained a low MAPE of 1.29%, proving its ability to adapt to structural shifts in energy demand.
Discussion and Conclusion
The results of this study highlight that "more data" is not always better; including all available variables often degrades model performance due to noise. By selecting a refined set of 5–7 key features, the LASSO-QRNN method achieves higher precision and better generalization.
Furthermore, the transition from point forecasting to probability density forecasting represents a significant advancement for energy planning. The generated PDFs consistently placed the actual consumption values near the peak probability (the mode), whereas traditional models often relegated real-world outcomes to the unlikely "tails" of their distributions. This research provides a powerful framework for energy planners to not just guess the future, but to quantify and manage the risks associated with it.
Summary updated