A new solar tracking sensor based on image recognition is proposed and designed to solve the problem of low accuracy of photoelectric tracking in photovoltaic power generation. The sensor can directly output its angular deviation from the sun, and its mechanical structure and working principle are analysed in detail. We use a high-precision camera to collect the image of the two slots on the projector surface and use the Hough transform to identify the image of the light seam. After obtaining the linear equation for the two slots, the coordinate of the intersection point is found, and the calculation of the solar altitude and azimuth can be realized. We have improved the Hough transform scheme by using the skeleton image of the slots instead of the edge image. The improvement of the scheme has been proved to effectively improve the detection accuracy. A calibration test board is used to test the sensor and experimental results show that the scheme can achieve the measurement of azimuth and altitude with the accuracy of be 0.05°, which can meet the detection accuracy requirements of the solar tracking in photovoltaic power generation and many other photoelectric tracking implementations.

At present, the back-propagation (BP) network algorithm widely used in the short-term output prediction of photovoltaic power stations has the disadvantage of ignoring meteorological factors and weather conditions in the input. The existing traditional BP prediction model lacks a variety of numerical optimization algorithms, such that the prediction error is large. The back-propagation (BP) neural network is easy to fall into local optimization thus reducing the prediction accuracy in photovoltaic power prediction. In order to solve this problem, an improved grey wolf optimization (GWO) algorithm is proposed to optimize the photovoltaic power prediction model of the BP neural network. So, an improved grey wolf optimization algorithm optimized BP neural network for a photovoltaic (PV) power prediction model is proposed. Dynamic weight strategy, tent mapping and particle swarm optimization (PSO) are introduced in the standard grey wolf optimization (GWO) to construct the PSO–GWO model. The relative error of the PSO–GWO–BP model predicted data is less than that of the BP model predicted data. The average relative error of PSO–GWO–BP and GWO–BP models is smaller, the average relative error of PSO–GWO–BP model is the smallest, and the prediction stability of the PSO–GWO–BP model is the best. The model stability and prediction accuracy of PSO–GWO–BP are better than those of GWO–BP and BP.