A novel transformer health state direct prediction method based on knowledge and data fusion-driven model

Predicting the future health state of a transformer can offer early warning of latent defects and faults within the transformer, thereby facilitating the formulation of power outage maintenance plans and power dispatch strategies. However, existing prediction methods based on the structure of ‘splicing prediction and diagnosis method’ suffer from limitations such as inability to achieve global optimality, error accumulation, and low prediction accuracy. To fill this gap, a novel direct prediction method of a transformer state based on knowledge and data fusion-driven model (K&DFDM) is proposed in this paper. Firstly, a state quantity data space is constructed to comprehensively reflect the changes in the health state of the transformer over time, encompassing online monitoring, offline testing, evaluation results, and actual operation data. After that, correlation knowledge between state quantities, fault diagnosis mechanism knowledge, current diagnosis experience knowledge, and uncertain fuzzy knowledge are extracted separately. The actual fault mechanism, existing expert experience, and other knowledge in the diagnosis process are quantified. Then, the attention model is subsequently optimised, leveraging quantitative knowledge to effectively constrain and guide the data prediction process. Incorporating fault diagnosis mechanism knowledge into the data prediction model enables the achievement of global optimisation in both diagnosis and prediction. The integration of traditional expert experience knowledge and the correlation knowledge between state quantities serves as constraints during the process of attaining the global optimum. The verification results, comprising 327 cases, demonstrate that K&DFDM effectively addresses the issue of error superposition encountered by existing state prediction methods, leading to a direct state prediction accuracy of 96.33%.