A Hybrid Autoencoder-Transformer Model for Robust Day-Ahead Electricity Price Forecasting under Extreme Conditions
Boyan Tang 1 2 , Xuanhao Ren 1 , Peng Xiao 1 , Shunbo Lei 3 , Xiaorong Sun 4 , Jianghua Wu 2 · 2025
Read A Hybrid Autoencoder-Transformer Model for Robust Day-Ahead Electricity Price Forecasting under Extreme Conditions by Boyan Tang 1 2 , Xuanhao Ren 1 , Peng Xiao 1 , Shunbo Lei 3 , Xiaorong Sun 4 , Jianghua Wu 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-06898.txt.
Introduction
Boyan Tang and five co-authors forecast day-ahead electricity prices, a task that matters for grid operation and market decisions. Prices spike unpredictably during extreme weather and holidays, and those anomalies defeat many current methods. The team combines two models: a distilled attention transformer that weighs key stretches of past data, and an autoencoder-based model that detects unusual patterns.
The transformer captures both long-term trends and short-term fluctuations through its self-attention weights. The autoencoder part, trained without supervision, isolates anomalies caused by events such as heavy rain, heat waves, or festivals, so they distort the forecast less. Tests on market data from California and Shandong Province showed better accuracy, robustness, and efficiency than leading methods, suggesting value for grid resilience.
Catalog record
| CREATOR | Boyan Tang 1 2 , Xuanhao Ren 1 , Peng Xiao 1 , Shunbo Lei 3 , Xiaorong Sun 4 , Jianghua Wu 2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-06898
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Hybrid Autoencoder-Transformer Model for Robust Day-Ahead Electricity Price Forecasting under Extreme Conditions Thanks: This work was supported by the Shenzhen Research Institute of Big Data (under the internal project No. J00220240006). ( Corresponding author: Jianghua Wu ) Boyan Tang 1 2 , Xuanhao Ren 1 , Peng Xiao 1 , Shunbo Lei 3 , Xiaorong Sun 4 , Jianghua Wu 2 Abstract Accurate day-ahead electricity price forecasting (DAEPF) is critical for the efficient operation of power systems, but extreme condition and market anomalies pose significant challenges to existing forecasting methods. To overcome these challenges, this paper proposes a novel hybrid deep learning framework that integrates a Distilled Attention Transformer (DAT) model and an Autoencoder Self-regression Model (ASM). The DAT leverages a self-attention mechanism to dynamically assign higher weights to critical segments of historical data, effectively capturing both long-term trends and short-term fluctuations. Concurrently, the ASM employs unsupervised learning to detect and isolate anomalous patterns induced by extreme conditions, such as heavy rain, heat waves, or human festivals. Experiments on datasets sampled from California and Shandong Province demonstrate that our framework significantly outperforms state-of-the-art methods in prediction accuracy, robustness, and computational efficiency. Our framework thus holds promise for enhancing grid resilience and optimizing market operations in future power systems. 1 1 footnotetext: For the emotional support function f(HalcyonYang) that always converges to happiness. (Corresponding author:TangBoyan) I Introduction Day-ahead electricity price forecasting (DAEPF) is vital to modern power system operations, providing important information for generators, market operators, and consumers. An accurate forecast helps generators schedule output, market operators formulate dispatch policies, and consumers plan energy usage and expenditure. In contrast, inaccurate forecasting threatens the stability of the whole market by triggering extreme price spikes or even negative electricity prices. Such danger has increased as electricity market regulations have increasingly been relaxed, allowing price signals to respond more strongly to variations in supply, demand, and extraneous factors. Recent events, e.g., record-breaking winter storms in parts of the United States and negative pricing events in Shandong province, China, have shown that extreme condition and other exceptional circumstances can increase price volatility, making improved forecasting capability a pressing necessity. DAEPF has been extensively studied using both statistical and machine learning approaches. Traditional statistical methods, such as Autoregressive (AR) [ 1 ] , Autoregressive Integrated Moving Average (ARIMA) [ 2 , 3 , 4 , 5 ] , and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) [ 6 ] , describe past patterns through time-series analysis. Although these methods effectively describe conventional price behavior, they typically depend on stationarity and normality assumptions that often fail under extreme market conditions. Moreover, they lack the flexibility to adapt to evolving dynamics, e.g., sudden policy changes or seasonal effects. Extensions like regime-switching models [ 7 ] and wavelet-based decompositions [ 8 ] have improved peak price accuracy. They, however, still perform inadequately when dealing with data exhibiting price spikes, heavy tails, and other complex behaviors. Encouraged by the success of neural networks in capturing non-linear relationships, machine learning methods began to gain popularity in electricity price forecasting in the 1990s. Early efforts were based on feedforward neural networks, e.g., multilayer perceptrons (MLP), and backpropagation-based models [ 9 ] . Subsequent research introduced more advanced network architectures, including convolutional neural networks (CNN) [ 10 ] and recurrent architectures such as long short-term memory (LSTM) [ 11 ] and gated recurrent units (GRU) [ 12 ] . These methods better capture temporal dependencies in electricity prices and tend to perform better than traditional time-series models. Ensemble methods, which combine multiple learning algorithms, have also been found to be effective in improving forecasting performance [ 13 , 14 ] . However, deep networks often become inefficient for capturing very long-term dependencies due to issues like exploding or vanishing gradients and high computational demands, rendering them less suitable for 24-hour day-ahead predictions or volatile market conditions. Consequently, researchers have recently adopted attention mechanisms and encoder-decoder architectures for DAEPF [ 15 , 16 ] . Such methods concentrate computational effort on the most informative parts of a sequence and better capture both local variations and overall market trends. Nevertheless, deep encoder-decoder models may still suffer from inefficiencies when processing large datasets or balancing short-term volatility with long-term policy impacts. To this end, this paper introduces a novel DAEPF framework built upon the Informer architecture [ 17 ] , which incorporates the following two key innovations: 1. Distilled Attention Transformer (DAT). Our method introduces a DAT model. DAT automatically identifies and emphasizes the most influential segments of historical data, enabling it to balance long-term trends with short-term fluctuations effectively. 2. Autoencoder Self-regression Model (ASM). To further improve the robustness under extreme conditions, our method integrates an Autoencoder self-regression model (ASM). ASM automatically recognizes and isolates atypical data patterns, such as those caused by heavy rain, heat waves, holidays, or massive human activities. The incorporation of the above two innovations contributes to superior DAEPF performance for our method. In particular, DAT model dynamically assigns higher weights to critical time steps, e.g., those influenced by sudden weather changes or unexpected events, thereby enhancing the model’s sensitivity to short-term anomalies. Concurrently, the ASM effectively detects and isolates these atypical patterns, ensuring that extreme events are modeled separately and do not distort the overall trend. This approach facilitates a balanced capture of long-term and short-term dependencies. Moreover, the streamlined computation inherent in the DAT model reduces the computational burden during real-time forecasting, while the modular architecture of the ASM decouples abnormal data processing from the primary forecasting task, thereby enhancing overall computational efficiency and responsiveness. We tested our model through experiments on datasets sampled from the California and Shandong Province markets, confirming that our framework consistently outperforms state-of-the-art baselines in terms of accuracy, responsiveness, and computational efficiency. The rest of this paper is organized as follows. Section II defines and formulates the DAEPF problem. Section III outlines the entire process for solving the DAEPF problem with our proposed framework. Section IV provides a detailed description of our proposed architecture, including its DAT model and ASM process. Section V presents experimental results on several datasets, including ablation studies and comparisons with state-of-the-art algorithms. Section VI discusses key findings and practical implications for DAEPF including conclusion and future works. II Problem Definition This section presents the definition and the formulation of DAEPF problem and introduces two corresponding forecasting strategies, including single-step and iterative multi-step prediction strategies. The latter recursively utilizes previous predictions as inputs for forecasting subsequent prices. To formally define the DAEPF problem, let X denote the historical electricity price, and is given as: X=\{x_{1},x_{2},\dots,x_{T}\}\ (1) where x_{t}\in\mathbb{R}^{d} represents the electricity price and related features (e.g., weather, load, etc.) at time step t . The goal of DAEPF problem is to predict the unknown electricity price Y for a future time period H : Y=\{y_{T+1},y_{T+2},\dots,y_{T+H}\} (2) where y_{T+h}\in\mathbb{R} denotes the electricity price at time step T+h . However, particularly, under extreme conditions, electricity price data may exhibit significant anomalous fluctuations, thus requiring a robust forecasting framework to handle both normal and extreme conditions. In some cases, the forecasting task may be performed iteratively, where the predicted result at the current time step is used as input to the next time step