In response to the issue of inaccurate innovation vectors and unknown measurement noise covariance matrices in the traditional unscented particle filter (UPF), a forecasting-aided state estimation (FASE) method for active distribution network is proposed, which integrates the improved Att-LSTNet and UPF. First, the key parameters of support vector regression (SVR) are optimized using a gravitational search algorithm (GSA), and a GSA-SVR model is established using historical data. This model is then introduced into the output layer of the Att-LSTNet model to create an enhanced forecasting model. Subsequently, the innovation vectors from UPF are used to train this model, and the isolation forest algorithm and box-plot method are employed to monitor and correct the original innovation vectors. Finally, in the case of unknown measurement noise covariance matrices, the corrected innovation vectors and UPF are combined to calculate the unknown measurement noise covariance matrices and perform state estimation. Case study results on the IEEE33-bus and IEEE118-bus test systems demonstrate the superiority of the proposed method in terms of estimation accuracy, generalizability, and robustness.