Accurate monthly temperature and precipitation forecasting in Guizhou Province is challenging because of its complex karst terrain and substantial spatiotemporal variability in meteorological conditions. We propose a forecasting framework that integrates Seasonal-Trend Decomposition using Loess (STL), Variational Mode Decomposition (VMD), bidirectional Long Short-Term Memory (BiLSTM) encoding, dynamic graph learning, graph convolution, and spatiotemporal cross-attention. STL and VMD provide multiscale temporal features, which are combined with meteorological, temporal, and station-related variables for spatiotemporal representation learning. An auxiliary high-precipitation loss is introduced to improve the prediction of high-precipitation events. Experiments were conducted using monthly observations from 38 meteorological stations in Guizhou Province during 1961-2011 under a chronological, leakage-free evaluation protocol. The proposed model achieved an RMSE of 1.759 °C for temperature and 57.550 mm for precipitation. Although it did not obtain the lowest aggregate precipitation error among the evaluated baselines, it achieved an upper-tail precipitation RMSE of 93.096 mm and MAE of 73.053 mm lower than those of monthly climatology, STGCN, and Graph WaveNet. Moving-block bootstrap analysis further supported the reductions in upper-tail RMSE and MAE relative to Graph WaveNet. These results indicate that the primary advantage of the proposed framework lies in forecasting high-precipitation magnitudes rather than minimizing aggregate precipitation error alone.