Local tube-through cables are the bottleneck locations of urban power grid current carrying capacity, the temperature monitoring of which is crucial. A real-time temperature inversion model based on LassoNet embedding and improved BP neural network is proposed to address the problems of low accuracy of current temperature measurement methods, inability to determine the optimal combination of temperature measurement points for different situations, and unsuitable input selection methods for "black box" networks. Firstly, the LassoNet network is used to autonomously quantify and select the optimal combination of temperature measurement points suitable for local tube-through cables neural networks; Subsequently, iCircle mapping, inertia weight concept, and Levy flight hybrid strategy are introduced to improve the initial distribution, search strategy, and iterative method of sparrow search algorithm (SSA) to enhance global optimization performance. The improved SSA is used to optimize the parameters of BP neural network, achieving fast and high-precision inversion of hot spot temperature under multiple operating conditions. A finite element simulation model of YJLW03-64/110 kV cable local conduit is established, and the accuracy of the model is verified by comparing it with IEC standards. Then, a hot spot temperature sample dataset is constructed under different load types. Based on this dataset, the proposed method is compared and analyzed with five typical inverse algorithms. At the same time, in order to verify the transferability of the algorithm, temperature inversion of 220 kV local tube-through cables and 110 kV cable joints is tested. The results show that the proposed inversion method can control the error within 1.5 ℃, has a fast convergence speed, can systematically select temperature measurement points, and has higher accuracy and robustness.