Price-preserving predictive energy management for blockchain mining in smart cities using a digital twin guided LSTM forecasting model

To reduce operating costs and minimize energy waste while maintaining service quality, smart cities require coordinated energy management across various interacting energy subsystems. This paper proposes a Cooperative Energy Framework (CEF) for the predictive management of energy used in blockchain mining, recognizing that demand variability can affect urban scheduling and pricing. Without utilizing price forecasting or shaping techniques, the predictive scheduling step aims to minimize the overall energy costs across subsystems, taking into account these hourly tariffs. Also, the proposed design integrates a Long Short-Term Memory (LSTM) forecasting model to predict short-term electricity requirements, along with a Digital Twin (DT) layer that represents miner telemetry. To ensure auditable data exchange and maintain record integrity among stakeholders, a blockchain-based data layer is utilized. Within operational limits, a cross-system optimization module allocates energy to meet mining demands while ensuring the functionality of other urban subsystems is preserved. The proposed method demonstrates a cost reduction of approximately 5.3% in the scenarios assessed, compared to a Traditional Model baseline that lacks cross-system cooperative coordination. The LSTM model also achieves a 30% decrease in Mean Squared Error (MSE) for forecasting compared to the Support Vector Regressor (SVR) baseline, and a 15% reduction compared to the Gated Recurrent Unit (GRU) baseline. Additionally, Mean Absolute Error (MAE) is utilized to evaluate performance. Also, we utilize Mean Absolute Error (MAE) and Mean Squared Error (MSE) to evaluate the miner-load forecasting module. A randomized split is employed to divide the dataset into subsets for training, validation, and testing. To prevent data leakage, feature normalization is applied only to the training subset before being utilized on the validation and test subsets. We compare several forecasting methods, including Gated Recurrent Unit (GRU), Support Vector Regression (SVR), Gradient Boosting Decision Trees (GBDT), Temporal Convolutional Network (TCN), and a Transformer-based forecaster, against Long Short-Term Memory (LSTM). The best model checkpoint is selected for test evaluation, and deep learning models are trained using early stopping based on validation loss. Overall, these findings indicate that combining constraint-aware coordination with DT-guided forecasting can enhance the efficiency and stability of urban energy operations in response to fluctuating mining demand.