Real-life water quality monitoring applications such as aquaculture domains and water resource management need long range multi-step prediction for disaster control. However, prediction accuracy usually degrades gradually as the prediction target timepoint is further away from the current timepoint. To address this, recent water quality forecasting methods mostly rely on complex deep learning models. In this paper, we propose a simple time-variant iterative ensembling method that strives to significantly improve the performance of a given arbitrary long range multi-step time series predictor for water quality data with minimal increase in computational cost. With the given predictor, our proposed method iteratively uses ensembles of predicted values for preceding steps to improve the prediction accuracy for the succeeding steps. The iterative ensembling operation is performed on the trained model and only at the inference stage, and so does not need any further computing-intensive training for the performance improvement. We experimentally show that the proposed method is effective with 7 predictors and 9 water quality datasets of various types, and it outperforms the state-of-the-art results in those datasets by around 2%–29% in mean absolute error (MAE), root mean squared error (RMSE) and mean absolute percentage error (MAPE) metrics. Similar improvement has also been found in two other metrics such as normalized Nash–Sutcliffe model efficiency coefficient (NNSE) metric and Taylor diagram plot. Overall, the proposed iterative ensembling is a promising approach for multi-step long range water quality prediction for high-frequency water quality monitoring systems.