To address limitations in the engineering application of neural network based maximum power point tracking(MPPT) algorithms, an improved lightweight neural network MPPT algorithm was proposed. The complexity and memory usage of the neural network were reduced through a knowledge distillation compression algorithm, and a lightweight model was obtained. The inherent theoretical error of model predictions was corrected using an optimized variable step-size perturb and observe method. In the initial stage, the neural network predicted the voltage range of the maximum power point. In the later stage, disturbance observation progressively refined this range until it converged at the maximum power point. A simulation model was developed in MATLAB/Simulink, and a physical model was constructed for comparative experiments. Results indicate that the proposed algorithm achieves higher tracking efficiency, improved ripple voltage suppression, and lower resource consumption rate in embedded devices.