A novel sensor layout method and deep learning model: leakage source term estimation for natural gas station sites

Accurate and efficient source term estimation (STE) is essential for ensuring the safe operation of natural gas facilities in the event of hazardous gas leakage. A well-designed sensor deployment, as the foundation of STE, directly affects the accuracy of source term estimation methods. Existing sensor placement methods are often limited to specific operating conditions and rely on prior information about the leakage source, making them impractical for real-world STE applications. Furthermore, traditional STE methods suffer from poor real-time performance and are difficult to adapt to dynamic leakage scenarios. This paper proposes a novel sensor optimization deployment method based on information entropy theory, designed to be applicable under multi-objective operating conditions. The proposed method employs a hybrid algorithm combining genetic algorithm (GA) and particle swarm optimization (PSO) to minimize the objective function, enabling optimal sensor placement under multiple wind direction conditions. Building on this foundation, the CEEMDAN algorithm is employed to denoise and reconstruct the sensor's noisy data. A fused model combining a bidirectional temporal convolutional network and a bidirectional gated recurrent unit (BiTCN-BiGRU) is then constructed to achieve high-precision estimation of the leak source term. Compared with traditional sensor deployment strategies, the optimized configuration improves the accuracy of source location estimation in STE methods by 24.5%, while reducing the estimation error of leak intensity by 87.3%. In terms of the STE model, the proposed CEEMDAN-BiTCN-BiGRU model achieves a leak intensity estimation accuracy of 99.90%, representing a 6.5% improvement in source location accuracy over the best-performing conventional model. These results demonstrate that the proposed sensor optimization approach and CEEMDAN-BiTCN-BiGRU model hold promise for applications in gas leak source term estimation. • A novel sensor optimization method for multi-objective working conditions is proposed. • Optimal sensor placement can be achieved without requiring prior knowledge of the leakage source. • A CEEMDAN-BiTCN-BiGRU model for source term estimation (STE) is proposed. • The accuracy of the CEEMDAN-BiTCN-BiGRU model can reach 99.90%.