Modeling of hydrological systems is generally based on time series, and instead of previously used linear statistical models, artificial neural networks are being used more and more recently. In the case of models for forecasting hydrological inflow, a review of the available literature revealed the problem of a delay in the model’s response after an actual event, which significantly reduces the usefulness of the model’s prediction. The same problem was confirmed in the model formed by classical methods for the Cetina river basin. The problem of timely forecasting in this dissertation is successfully solved by introducing the variable frequency of forecasted precipitation and by adjusting the calculation step. Forecasting accuracy is improved by forming an optimized adaptive neural model (OANM) based on dividing the model into specialized sub-models, which represent specially trained networks for forecasting at certain times of the year. Each submodel was formed according to the principles of a systematic approach to the formation of neural models for hydrological predictions. This means that the optimal input variables, the number of neurons of the hidden layer, the algorithm for training the network and the objective function were determined through specially formed experiments, and a set of optimal measures for evaluating the model was also determined. A static feedforward neural network, a multi-layer perceptron, was used to form all models, due to its proven property of being able to approximate a non-linear function. The MATLAB program was used for all research, as well as the final formation of OANM, and the data were processed in Microsoft Excel in addition to Matlab. The simulation results confirm the success of the proposed method in solving the problem of timely and reliable forecasting of hydrological inflow, which means that the proposed method successfully eliminates the delay of model prediction.