Dynamic update model of transmission and distribution production command knowledge base combining deep learning with operation and maintenance data

Production command and maintenance planning within the context of the Transmission and Distribution (T&D) network has to be supported in a reliable way, with data. The traditional transformer operation and maintenance practices are mainly based on rule-based analysis and expert judgement, which has limited predictive ability when operating in the sparse and noisy fault condition. To alleviate this, a model that is combining deep learning with historical data is proposed. Experiments are conducted on a 15k-record operational fault data set with measurements, states, fault logs and maintenance data created by T&D. The data has been preprocessed by cleaning up the data, handling missing values, and normalizing the data with Z-scores to ensure consistency and numerical stability before it is used for model training. With PCA, feature extraction is carried out, which means that the dimensionality is reduced, while the critical operational information is recalled. A knowledge graph is built based on the semantic representations of complex relationships between equipment, faults and operational parameters, allowing for structured knowledge management and ongoing updates. For temporal learning, a Dynamic Orca Prediction-optimized Vanilla LSTM (DOP-Vanilla LSTM) is employed. The Vanilla LSTM captures long-term dependencies from high-dimensional and limited fault data. The DOP algorithm enhances parameter optimization, learning stability, and convergence efficiency. A dynamic update mechanism incrementally integrates newly acquired Operation and Maintenance (O&M) data and prediction outputs, improving adaptability to evolving grid conditions. Experimental evaluation is conducted using Python-based tools, including TensorFlow and Scikit-learn. Results demonstrate that the proposed approach achieves prediction accuracy in the range of 95%–98%, outperforming existing methods such as RF and conventional deep learning baselines. The framework enhances fault diagnosis accuracy, production command reliability, and intelligent decision-making for modern T&D systems.