Automated Neural Network Auditing (ANNA) for Energy Market Settlements: A Machine and Reinforcement Learning-Based Approach

Satish Mahadevan Srinivasan, John Raymond Rivadeneyra

Keywords: Energy Market Settlements, Automated Auditing, Neural Networks, Anomaly Detection, Reinforcement Learning, Machine Learning.

Abstract: Energy market settlements in ISOs such as CAISO and MISO have grown increasingly complex with the expansion of EIMs and automated transaction processes. Traditional rule-based, manual auditing is inefficient, error-prone, and unable to manage the massive data volumes these markets generate. To address this, we propose Automated Neural Network Auditing (ANNA), a hybrid machine and reinforcement learning framework for scalable anomaly detection. ANNA applies neural networks for anomaly identification and reinforcement learning for adaptive, real-time auditing, incorporating statistical methods, K-nearest Neighbors (KNN), and Deep Q-Networks (DQN). Using three years of CAISO settlement data covering 32 charge codes and over 63 million transactions, ANNA is evaluated against traditional statistical and clustering models on precision, recall, efficiency, and false negative reduction. Preliminary results show ANNA outperforms conventional approaches by uncovering hidden relationships in settlement data, enabling more accurate, compliant, and efficient audits.

Primary Area: reinforcement learning

Submission Number: 10025