The patients with type 1 diabetes (T1D) lack natural insulin secretion and need daily insulin injections to keep their blood glucose (BG) within an appropriate range. Artificial pancreas (AP) system is a promising therapeutic method to solve that problem, which comprised three parts, namely, a continuous glucose monitor (CGM), an insulin pump, and an intelligent controller. The intelligent controller plays a key role because it calculates the appropriate insulin amount for the insulin pump. We have developed an AP controller based on generalized predictive control (GPC). But it only uses linear equations and could not imitate the complex blood glucose-insulin dynamics. The nonlinear autoregressive moving average (NARMA-L2) is a simple neural model but an effective way to represent nonlinear systems. Here, a NARMA-L2 controller was proposed for AP system. Tests results showed that it effectively regulated the BG of 9 in-silico patients with an average percentage of time within an appropriate range of 77.43%. However, two subjects (No. 05 and 06) still had a high hyperglycemia risk and other two subjects (No. 02 and 07) had a high hypoglycemia risk. Thus, subsequent work is needed and the NARMA-L2 controller should to be further optimized to better prevent the hyperglycemia and hypoglycemia events.