Laser methane sensor has obvious advantages of anti-poisoning, anti-interference, and long service life. It can be used for real-time online monitoring of natural gas leakage in complex environments. However, the laser methane sensor is easily affected by temperature, resulting in a large difference between the actual measured CH₄ concentration and the actual value. Common temperature compensation algorithms include polynomial fitting method and empirical formula method. These two temperature compensation methods have a good effect on temperature compensation under the influence of single factor. However, the influence factors of temperature on the laser methane sensor include gas molecules, optical elements and circuit elements. Therefore, in the actual quality application, there is still a large error between the corrected concentration value and the true value.In this paper, a temperature compensation model is established by using the deep learning method. Its prediction accuracy mainly depends on the network model structure and large-scale training samples. In order to effectively improve the detection accuracy of the laser methane sensor in a wide temperature application environment, combined with industrialization, a large-scale laser methane sensor high and low temperature detection sample data set was established, and the model effect was further improved through big data training.Based on the model framework, an ISSA-BP algorithm with global optimization capability is proposed. Firstly, a quasi-reflective learning strategy is used to initialize the sparrow population to improve the efficiency of iterative optimization. Secondly, we use the strategy of searching for prey in CSA to improve the location update of explorers in SSA, so that the algorithm has the ability to jump out of local optimization. At the same time, Levy flight strategy is introduced to improve the anti-predator position update and enhance its global search ability. Finally, the artificial rabbit disturbance strategy is used to update the sparrow individuals to further reduce the probability of the algorithm falling into the local optimum. By using the standard sparrow search algorithm, particle swarm optimization algorithm and grey wolf optimization algorithm to test unimodal function and multimodal function, the advantages of ISSA in terms of convergence accuracy and speed, global search and local development capability are verified.In terms of data, the training effect of the temperature compensation model is improved and the prediction error of the model is reduced by establishing a large-scale experimental data set of sensors with different temperatures and concentrations. In the temperature range of -20 ℃~65 ℃, 15 800 groups of sensor measurement data were used to carry out comparative experiments on BP, PSO-BP, SSA-BP and ISSA-BP temperature compensation models. The results show that the maximum relative error of the predicted value of temperature compensation model based on ISSA-BP neural network is only 0.52%, which is 7.70%, 2.46%, and 0.74% less than that of BP, PSO-BP, SSA-BP models respectively. When the temperature changes from -20 ℃~65 ℃, the predicted value of concentration still fluctuates in a small range. The Average Absolute Percentage Error (MAPE) of BP neural network, PSO-BP neural network, SSA-BP neural network and ISSA-BP neural network for predicting the test sample of 0.5% standard concentration methane gas is 0.038 6%, 0.014 6%, 0.005 8%, and 0.002 7%, respectively. Compared with other models, the values of MAE, MAPE, RMSE and RE of ISSA-BP neural network model are smaller, which indicates that ISSA-BP temperature compensation model has higher accuracy and better stability.The research results show that the algorithm in this paper can greatly improve the detection accuracy of the laser methane sensor in a wide temperature application environment, and is of great significance in improving the environmental applicability of the laser methane sensor.