Multi-objective predictive maintenance optimization of electric energy meters based on LSTM-XGBoost and modified firefly algorithm

Abstract Due to the rapid development of modern smart grid and Internet of Things technology, the operational and maintenance demands on smart meters, the core metering equipment of the power system, are becoming increasingly significant. Traditional maintenance methods waste resources and are inefficient. Therefore, a new maintenance paradigm combining stable, efficient prediction technology and intelligent, decision-making optimization is urgently needed. This paper proposes a multi-objective predictive maintenance optimization framework based on a long short-term memory (LSTM)-extreme gradient boosting (XGBoost) hybrid model and an improved firefly algorithm (IFA). The framework addresses three core challenges in smart meter operation and maintenance: complex failure modes, limited resources, and conflicting objectives. The main innovations of this study are: 1. A spatiotemporal hybrid prediction model for meter failures is proposed for the first time. This model combines bidirectional LSTMs to capture long-term parameter dependencies, such as voltage and current, and introduces an enhanced XGBoost with a dynamic threshold Huber loss function (DTHL). This reduces error by 25 % in voltage mutation scenarios. 2. A maintenance decision optimization model is designed based on an improved firefly algorithm. The model suggests a maintenance-specific brightness index and a cosine annealing step strategy. The convergence speed is 3.2 times faster than NSGA-II. 3. Support edge-cloud collaborative computing with a prediction response time of less than 200 ms. Using real data from the Gansu Electric Power Company from 2020 to 2024 covering 9,823 devices, the LSTM-XGBoost model performs well in predicting faults. Under budget and manpower constraints, IFA’s cost savings and risk reduction rates are 8.2 % and 9.5 % higher than NSGA-II, respectively, and resource utilization is 89.7 %. SHAP analysis shows that temperature, voltage fluctuations and communication packet loss rate are key fault characteristics.