Given that environmental factors such as solar radiation, wind speed, and temperature exert significant influences on the fluctuation of photovoltaic (PV) power generation, these fluctuations pose substantial challenges to the secure operation and dispatching of power systems with large-scale PV integration. Therefore, accurate PV power forecasting is essential for maintaining grid stability. A hybrid forecasting model, CW-MIDBO-ELM, was proposed. The model integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), wavelet analysis, the multi-strategy improved dung beetle optimization algorithm (MIDBO), and the extreme learning machine (ELM). CEEMDAN combined with wavelet denoising was employed to decompose and reconstruct key meteorological features affecting PV output, thereby mitigating the non-stationarity of the input sequences. Furthermore, the MIDBO algorithm was used to optimize the input-layer weights and hidden-layer bias parameters of the ELM, enabling improved network stability and enhanced prediction accuracy. Experimental results demonstrate that the proposed CW-MIDBO-ELM model achieves markedly higher forecasting performance compared with benchmark methods.