The data used for intraday wind power forecasting (WPF) are often collected in non-stationary environments. Accuracy can be substantially reduced due to concept drift caused by significant changes in operational conditions of wind farms or in the probability distribution of samples. An adaptive intraday WPF model considering multi-timescale concept drift detection is proposed to address the above challenges. First, a spatio-temporal forecasting model is constructed to achieve intraday WPF. Then, an integrated mask-reconstruction representation learning pretraining strategy is employed to extract hidden representations of input historical wind power measurements and numerical weather prediction data. The degree of concept drift in the sample stream is quantified by measuring the cosine similarity between current and historical hidden representations. Finally, two concept drift detection modules with different time-scales are employed to guide the model in performing multi-stage adaptive update, enabling it to accommodate varying degrees of concept drift and achieve a balance between accuracy and flexibility during the detection process. Case studies based on 3 real wind farm clusters demonstrate the proposed method's superior forecasting accuracy and computational efficiency.