To address the issue of low prediction accuracy caused by the unstable and fluctuating characteristics of photovoltaic (PV) power generation, a short-term PV power forecasting method is proposed. The method integrates quadratic mode decomposition reconstruction (QMDR), a bidirectional temporal convolutional network (BiTCN) and bidirectional gated recirculation unit (BiGRU) combined model, and a multi-strategy improved sand cat swarm optimization algorithm (MSCSO). First, meteorological features are selected as model inputs using the Spearman correlation coefficient, and fuzzy C-mean clustering method is applied for similar-day classification. Next, the PV power series are decomposed by improved complete ensemble empirical modal decomposition and variational modal decomposition, and the components are reconstructed by sample entropy. Finally, a combined prediction model of BiTCN-BiGRU is established, with the parameters of the model optimized by MSCSO. The final PV power prediction is obtained by superimposing the forecasts of each constructed component. Comparative analyses under different weather conditions and across different regions verify that the proposed model has higher prediction accuracy and better adaptability than existing approaches.