Accurate wind power forecasting is vital for grid stability, but sensor data are frequently compromised by environmental degradation and hardware anomalies. To address this limitation, we developed an intelligent forecasting model that enhances sensor reliability by integrating K-dimensional tree-accelerated density-based spatial clustering of applications with noise (KD-DBSCAN) with a bidirectional long short-term memory (BiLSTM) network optimized by a rime optimization algorithm (RIME). KD-DBSCAN reduces spatial search complexity to O(NlogN) and purifies multi-dimensional sensor data. The purified data feed into the RIME-BiLSTM architecture. Tested on two commercial supervisory control and data acquisition (SCADA) datasets, KD-DBSCAN anomaly filtering improved the coefficient of determination (R2) by up to 48.4%, reducing forecasting errors by 39.6% (inland commercial wind farms) and 32.8% (coastal wind farms). The optimized RIME-BiLSTM showed an R2 of 0.9876 and reduced the mean absolute error by up to 48.9%. Beyond soft-sensing, isolating physical anomaly signatures, such as sensor drift and signal dropout, provides diagnostic metrics to guide the development of novel physical sensors, such as solid-state ultrasonic wind transducers with anti-icing hydrophobic coatings and self-calibrating microsensor modules. Despite these results, reliance on offline batch processing poses latency constraints for streaming SCADA. Therefore, an online incremental KD-DBSCAN framework must be developed to be paired with adaptive continual learning for real-time edge deployment.