City gas load forecasting is significant to the operation of city gas networks. In consideration of the periodicity and nonlinearity of gas load data and the shortcomings of a single model, a hybrid model of Echo State Network (ESN) and improved RBF Neural Network (RBFNN) is put forward. First of all, kernel Fisher linear discriminant is utilized for dimension reduction. Secondly, we adopt ESN to do a preliminary prediction. Then, differential evolution integrated with gradient descent by encoding is used to learn and optimize the structure and parameters of RBFNN. Last but not least, the produced result of ESN is the input of RBFNN. It is validated that the proposed model has a higher precision and convergence rate compared with the initial combinational model.