The advent of renewable energy has huge implications for the design and control of power grids. Due to increasing supply-side uncertainty, traditional reliability constraints such as strict bounds on current, voltage and temperature in a transmission line have to be replaced by chance constraints which are computationally hard. In this paper we use large deviations techniques to study the probability of current and temperature overloads in a DC network with stochastic power injections, and develop corresponding safe capacity regions. In particular, we characterize the set of admissible power injections such that the probability of overloading of any line over a given time interval stays below a fixed target. We show how enforcing (stochastic) constraints on temperature, rather than on current, results in a less conservative approach and can thus lead to capacity gains in power grids.