Load Forecasting of Power System Based on Deep Neural Network and Digital-Twin

With the accelerated development of the global energy Internet and the integration of green and clean energy, the future power grid has posed more strict demands for the accuracy and stability of power load prediction. This paper proposes a novel deep neural network model, TCN-LSTM, that integrates Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM), and is enhanced by an adaptive attention mechanism and a hybrid loss function (MSE-CE). The main contributions of the research are Architectural Innovation, Training Optimization, and Empirical Validation. A dual-module design leverages TCN to capture long-term dependencies and LSTM to model temporal dynamics, while the attention mechanism dynamically prioritizes critical features, significantly improving feature utilization. Furthermore, the research combines the designed prediction model with a digital twin to construct a dynamic simulation environment for the power system, enabling real-time mapping and predictive feedback on the power grid’s operating status. This enables real-time updates and dynamic optimization of the prediction model, enhancing the adaptive capability of load forecasting. The comparison index data show that, in short-term load forecasting, the average absolute error, average absolute percentage error, and root mean square error of the proposed algorithm in the research are 52.1, 0.0059, and 70.9, respectively. Compared with other algorithms, they are reduced by an average of 30.03, 0.0028, and 26, respectively. In terms of long-term prediction results, the average absolute error, average absolute percentage error, and root mean square error of the proposed algorithm in the research are 57.2, 0.0081, and 80.9, respectively, which are on average reduced by 39.43, 0.0062, and 32.47 compared with other algorithms. These results validate the high accuracy and stability of the proposed method for load forecasting.