Modeling bitcoin network energy demand: Price-adjusted hybrid deep learning approach to complex time series forecasting

This study proposes a comprehensive deep learning framework to predict bitcoin energy usage by integrating advanced data preparation, feature engineering, and statistical analysis. The methodology utilizes a novel custom dataset combining the Cambridge Bitcoin Electricity Consumption Index (CBECI) and real-world industrial electricity prices from the United States Energy Information Administration (EIA), resulting in a robust, price-adjusted energy consumption time series. Stationarity is assessed using the Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests, while Seasonal-Trend decomposition using Loess smoothing (STL) and AutoRegressive Integrated Moving Average (Auto-ARIMA) guide detrending and differencing. Feature selection is performed via mutual information and Granger causality analysis to identify key attributes related to energy consumption. Four deep learning architectures, namely Long Short-Term Memory (LSTM), one dimensional Convolutional Neural (1D-CNN), hybrid CNN-LSTM, and a convolutional transformer-based model (ConvTransformer), are evaluated using a sliding window approach. Model performance is assessed with various metrics. Our results reveal that the ConvTransformer consistently outperforms other models, achieving superior accuracy and stability with the lowest mean absolute error (MAE = 1.463), root mean square error (RMSE = 1.892), and Huber Loss (1.038), as well as the highest coefficient of determination (R² = 99.355 %). These findings establish a robust framework for modeling complex, non-stationary, high-dimensional time series with irregular patterns, extending beyond bitcoin to broader energy forecasting in computational systems. The methodology contributes to sustainable blockchain development by enabling accurate energy consumption predictions essential for environmental impact assessment and policy formulation.