In this study, an artificial neural network (ANN) approach was applied for modeling the relation between environmental variables and daily change in the Cleanness Index (ΔCI, a measure of performance loss due to soiling) of photovoltaic modules in the field in Doha, Qatar. The daily ΔCI was examined among a number of three-dimensional intervals of the daily mean of the environmental variables (i.e., the intervals of two environmental variables were presented on x and y dimensions, and average values of daily ΔCI on the third dimension), in order to qualitatively establish the relations that might help to develop improved PV soiling prediction models. Then, an ANN-based model was set up to simulate the relationship between daily ΔCI and environmental variables and compared with a linear regression model, both models using the same input variables, including present day and previous day environmental conditions, and cumulative exposure time. Strong interactions were observed among environmental variables PM10, relative humidity (RH) and wind speed (WS) regarding their effect on the daily ΔCI. Overall, higher PM10 resulted in more negative daily ΔCI (i.e. the module became more soiled), and this effect was more visible at low WS and RH levels, but at high WS (>4ms−1) and high RH (>65%) levels, PM10 had no significant (p>0.05, two tailed t-test) effect on daily ΔCI. Mostly, WS and RH determined how much airborne dust accumulates on the module surfaces and thereby affects the output of the PV modules. Higher WS typically favored more positive daily ΔCI when RH was low, but at higher RH levels (>50%) daily ΔCI was more likely to be negative with increasing WS. In fact, high RH levels were related to negative daily ΔCI only at higher WS levels (>2ms−1); at lower WS levels RH had no significant effect on daily ΔCI. These effects were apparently due to the deposition-resuspension mechanisms of dust accumulation on the PV panel surfaces. The ANN model performed significantly better in predicting daily ΔCIas well as cumulative CI than the linear model in term of R2 values and statistical error indexes. The previous day environmental conditions had a significant effect on the modeling outcome. The inclusion of the wind gustiness and cumulative exposure time also considerably improved the model prediction capability. The advantage of the ANN-based model is its simplicity, ease of data fitting and no requirement of an accurate mathematical model.