In order to obtain more accurate rolling force of horizontal roller, the actual rolling parameters of Fujian Luoyuan Minguang Iron and Steel Rolling Mill were collected, and the calculation and preprocessing of relevant parameter were performed to construct a rolling force dataset of horizontal roller for H-beam containing multi-input features and multiple specifications. To effectively predict the rolling force of horizontal roller for H-beam, firstly, the outlier detection and feature selection were conducted by isolation forest algorithm and tree model, and the dataset was divided, using random forest model as the base model for training and validation. Next, the random forest model was optimized by northern goshawk optimization algorithm. Furthermore, the processed test set data of rolling force for H-beam horizontal roller was inputted, and the predicted rolling force values were output. In addition, the constructed model (NGO-RF) was compared with the unoptimized random forest model, support vector machine model, multi-layer perceptron model, convolutional neural network model, as well as support vector machine model and multi-layer perceptron neural network model optimized by northern goshawk optimization algorithm. The results show that the constructed model outperforms all models mentioned above in terms of performance prediction, and it has high accuracy and adaptation. Additionally, the rolling force of new H-beam 588 mm×300 mm×12 mm×20 mm specification products are predicted by using the constructed model. Comparing the predicted values of the model with the actual measured values, the average error is only 6.05%, further confirming that the constructed model can effectively predict the rolling force of H-beams horizontal roller.