To address the limited nonlinear feature extraction capability of day-ahead forecasting models and the difficulty of model learning due to complex power output fluctuations in intraday forecasting, a multi-time scale photovoltaic (PV) power forecasting method based on Kolmogorov-Arnold networks (KAN) and the autocorrelation features of error time series is proposed. First, in the day-ahead forecasting stage, a forecasting model using KAN as the basic building block is designed. A deep KAN architecture enhanced with residual connections is used to extract spatial features, while a multi-head attention mechanism is employed to extract temporal features, significantly improving the model’s ability to capture diverse climatic characteristics. Then, in the intraday forecasting stage, based on the day-ahead forecasting results, indirect prediction is performed by incorporating the autocorrelation features of the error time series. This approach significantly reduces the fluctuation range of the predicted sequence and lowers the learning difficulty of the model. Finally, experiments conducted using data provided by a PV power forecasting competition demonstrate that the proposed day-ahead model reduces the mean square error (MSE) by at least 3.8% compared with long-short-term memory (LSTM) and Transformer models. Compared with direct forecasting, the MSE of the forecasting results is reduced by 38.3%.