We focus in this paper on the transport of critical services in unlicensed spectrum, where stringent constraints on latency and reliability are to be met, in the context of Ultra-Reliable Low Latency Communication (URLLC). Since contention-based medium access performs poorly in the case of high traffic load, we propose a new transmission scheme where the transmitter can increase its transmission power when the delay of the packet approaches the delay constraint, increasing by that its chance of being decoded even in case of collision with other lower-power packets. We are however interested in minimizing the usage of high power transmissions, mainly to conserve energy for battery-powered devices and to limit the range of interference. Therefore, we define a transmission policy that makes use of a delay threshold after which the high-power transmission starts, and propose a new online-learning approach based on Multi-Armed Bandit (MAB) in order to identify the policy which achieves minimum energy consumption while guaranteeing reliability. However, we observe that the MAB converges slowly to the optimal policy because the loss event is rare in the load regime of interest. We then propose a model-aided learning approach where a simple analytical model helps estimating the longterm reliability resulting from an action and thus its reward. Our results show a significant enhancement of the convergence towards the optimal policy.