Optimization of energy forecasting anomalies using LSTM based time series analysis and secure federated learning

The global energy sector is prioritizing the development of AMI (Advanced-Metering Infrastructure) to enhance the smart grid sustainability, where energy load forecasting anomalies remain a critical challenge. Existing machine learning-based load forecasting models exhibit anomalies, such as underestimation and overestimation of energy demand, which contribute to errors in AI (Artificial Intelligence) hallucination (inaccurate generative forecasting). Moreover, cyber threat anomalies can emerge in federated learning and smart meter node-based load forecasting processes, including risks, such as information leakage and data breaches caused by cyber-attack. To overcome these challenges, this paper proposes a generative load forecasting approach by combining time-series comparative analysis with LSTM (Long-Short-Term-Memory) neural networks. The load forecasting program computes generative load forecasts at each smart meter node, which are encrypted using AES (Advanced-Encryption-Standard) cryptography. The encrypted data is aggregated into AMI server and a total load forecast for a substation-grid node is generated through federated learning aggregation method. In the proposed Secure-LSTM-FedAggSum process, the smart-meter energy data used that was provided by the UK's 'Low Carbon-London' project. The comparative analysis shows that the proposed model outperforms other approaches in mitigating load-forecasting anomalies and delivers more reliable, risk-free forecasting.