In this paper, we explore the possibility of learning heuristics for solving numeric planning tasks. In numeric planning, states may exhibit numeric variables and state transitions are defined by mathematical expressions over these variables. We propose two new learning architectures for numeric planning. Our first approach involves constructing a new graph kernel for graphs with both continuous and categorical attributes, and using it to generate features for numeric planning tasks. This method is quick to compute, transparent and explainable. Our second approach makes use of graph neural networks. Experimental results show that our proposed methods yield competitive coverage performance over several numeric planning benchmarks in comparison to domain-independent planners.