Affected by road condition and human factors, the practical traffic system can be considered as a complex nonlinear dynamical system and the traffic flow data have the properties of strong non-linearity, time-variety and susceptibility to random noises. To solve the problems of short-term traffic flow forecasting in complex environment, a prediction method based on the hybrid FWADE-ELM is proposed. The singular spectrum analysis(SSA) technique is used to filter the noise existing in the original traffic flow data and then the ELM neural network prediction model is trained with the preprocessed data. The structure and key parameters of ELM network are determined using the C-C algorithm after reconstructing phase space. A hybrid optimization method integrating the fireworks algorithm with the differential evolution algorithm is developed to improve the optimal performance of the basic algorithms. The ELM network forecasting model is built and its weights and biases (9-11-1 structure and 110 dimensions) are optimized using the proposed FWADE hybrid algorithm. The results of property testing and practical application show that this short-term traffic flow forecasting model has higher forecasting accuracy and better generalization ability, and the predicted values comply well with the actual values.