Power load forecasting combining deep learning models and improved CLPO algorithm

The accuracy of power load forecasting directly affects the stable operation of the power system and energy optimization configuration. In response to the complex nonlinear characteristics of load data and the tendency of existing methods to fall into local optima, this study selects time-series convolutional networks, Long Short-Term Memory (LSTM) networks, and bidirectional LSTM to propose a combined deep learning model through weighted reconstruction. On this basis, the parrot optimization algorithm is introduced for iteration and strategy optimization, and joint error correction is performed to propose a new power load forecasting model. The experimental results show that under the same conditions and with the same hyperparameters, the conventional gradient descent optimization method takes about 330 rounds to reach 0.85 and the highest is only about 0.89. While the proposed method reaches 0.85 in about 200 rounds and eventually approaches 0.95. The overall convergence is faster and the accuracy is higher. On the UK-NGED dataset, the long-term prediction mean absolute errors of Transformer-DGNN, EEMD-Attention and VMD-ResNet are 0.214, 0.238 and 0.257 respectively, and the research method is 0.181. Simultaneously, continuous learning-based parrot optimization and error correction model suppress weight drift and short-term bias during optimization, maintaining prediction stability under concept drift and non-steady load patterns. When prediction performance degrades, error-monitoring-based lightweight retraining rapidly restores model accuracy. Under extreme weather conditions, the root mean square error of the baseline method is 0.295-0.318, and that of the research method is 0.250. To confirm that the differences are not due to random fluctuations, a two-sample t-test is conducted on the root mean square error distributions from 10 independent experiments. After explicitly calculating the mean difference, variance, and degrees of freedom, the result yields p < 0.05, indicating that the reduction in model error under extreme meteorological conditions is statistically significant. Therefore, the model can maintain high accuracy prediction in variable load environments and has stronger adaptability in high penetration rates of new energy and extreme weather conditions. This study provides more accurate and reliable decision support for smart grid scheduling and optimization.