Accurate intraday forecasts of PhotoVoltaic (PV) system power outputs are crucial for improving energy distribution grid operations. We present a neural autoregressive model for such forecasts, building upon a physical, deterministic PV performance model. Our approach uses the physical model's output as covariates and addresses multiple PV sites with a single neural model. We introduce a truncated Gaussian output distribution novel in this context and use a scale-free approach with explicit modeling of seasonal effects. The interest of the approach is validated on real data.