The development of electricity spot markets necessitates more refined and accurate load forecasting capabilities to enable precise dispatch control and the creation of new trading products. Accurate load forecasting relies on high-quality historical load data, with complete load data serving as the cornerstone for both forecasting and transactions in electricity spot markets. However, historical load data at the distribution network or user level often suffers from anomalies and missing values. Data-driven methods have been widely adopted for anomaly detection due to their independence from prior expert knowledge and precise physical models. Nevertheless, single network architectures struggle to adapt to the diverse load characteristics of distribution networks or users, hindering the effective capture of anomaly patterns. This paper proposes a PLS-VAE-BiLSTM-based method for anomaly identification and correction in load data by combining the strengths of Variational Autoencoders (VAE) and Bidirectional Long Short-Term Memory Networks (BiLSTM). This method begins with data preprocessing, including normalization and preliminary missing value imputation based on Partial Least Squares (PLS). Subsequently, a hybrid VAE-BiLSTM model is constructed and trained on a loaded dataset incorporating influencing factors to learn the relationships between different data features. Anomalies are identified and corrected by calculating the deviation between the model’s reconstructed values and the actual values. Finally, validation on both public and private datasets demonstrates that the PLS-VAE-BiLSTM model achieves average performance metrics of 98.44% precision, 94% recall rate, and 96.05% F1 score. Compared with VAE-LSTM, PSO-PFCM, and WTRR models, the proposed method exhibits superior overall anomaly detection performance.

Accurate extraction of typical load curves (TLCs) is essential for electricity market trading, demand-side management, and optimal design of energy storage systems. However, conventional methods are highly sensitive to anomalous consumption days caused by equipment failures or maintenance, which can distort normal electricity consumption patterns. To address this issue, this paper proposes a two-stage unsupervised framework that integrates a deep sequence model with an anomaly detection algorithm for robust TLC extraction. First, a Transformer-based autoencoder is employed to learn complex temporal dependencies and intrinsic patterns from historical daily load data, extracting robust periodic features by reconstructing the input load sequences. Subsequently, the reconstruction error of each daily load curve is computed as an anomaly assessment metric. These reconstruction error features are then fed into an Isolation Forest algorithm to identify anomaly loads that significantly deviate from the learned normal patterns, without requiring predefined thresholds or labeled data. Validation using real-world commercial and industrial electricity consumption data demonstrates that the proposed method effectively filters out various anomalies (e.g., spikes, troughs, and shape distortions) that conventional methods fail to exclude. The extracted TLCs exhibit improved robustness and representativeness. Further case studies indicate that adopting purified TLCs to guide electricity procurement in market trading facilitates more scientific trading strategies and avoids increased electricity costs caused by distorted load patterns. In summary, the proposed Transformer-Isolation Forest hybrid framework provides an effective data-driven solution for robust TLC extraction. The resulting TLCs can be directly used to guide day-ahead market bidding, optimize power purchase contract decomposition, and assess user demand response potential.