Groundwater plays a vital role in securing water supply in arid and semi-arid regions. Most previous studies have forecast groundwater level as a function of hydrological parameters such as rainfall, temperature, and streamflow; however, limited access to such data in arid regions constrains the applicability of these approaches. This study presents a novel approach in which only the historical time series of groundwater level itself is used as input data, so that the groundwater level at each time step is forecast from observed values at previous time steps (with lags of up to five months). Forecasting models based on an artificial neural network (ANN) and a wavelet neural network (WNN) were developed and evaluated for forecasting the monthly average groundwater level of the Kerman plain over the period October 1987 to March 2024. Results showed that the WNN model, with a test-stage MSE of 0.0139 m², provided substantially higher accuracy and generalization than the ANN model (MSE = 0.421 m²). Among the three optimization algorithms applied (GA, ICA, and PSO), the imperialist competitive algorithm (ICA) achieved the best performance, with MSE = 9.328×10⁻⁵. These findings indicate that coupling discrete wavelet decomposition with an artificial neural network, using only historical groundwater-level data, is a practical and reliable approach for groundwater-level forecasting in arid and semi-arid regions with limited hydrological data.