1: \begin{abstract}

2: 	This paper analyzes the stability-related properties of Long Short-Term Memory (LSTM) networks and investigates their use as the model of the plant in the design of Model Predictive Controllers (MPC).

3: 	First, sufficient conditions guaranteeing the Input-to-State stability (ISS) and Incremental Input-to-State stability ($\delta$ISS) of LSTM are derived.

4: 	These properties are then exploited to design an observer with guaranteed convergence of the state estimate to the true one. 

5: 	Such observer is then embedded in a MPC scheme solving the tracking problem. The resulting closed-loop scheme is proved to be asymptotically stable.

6: 	The training algorithm and control scheme are tested numerically on the simulator of a \emph{pH} reactor, and the reported results confirm the effectiveness of the proposed approach.	

7: 	

8: 	\textit{Keywords.} Learning-based control; Nonlinear model predictive control; Output feedback predictive control; Long short-term memory neural networks; Machine learning

9: 	\end{abstract}

10: