Abstarct: To increase processor computational power, the designers of the microprocessors have built multicore processor on a chip. Adding to the number of cores causes an increase in the power density and subsequently leads to raising the processor temperature. To enhance the processor performance and prevent from burning, its temperature should be managed and controlled. In order to manage the temperature, reactive and proactive approaches have been introduced. Contrary to the reactive approach, proactive methods control the temperature before reaching the temperature threshold. One of the main challenges in proactive methods is the need for a model that predicts the processor temperature with high precision. Learning this model requires a dataset that includes a large variety of processor temperature variations. For this purpose, an algorithm is proposed to create a proper dataset. A number of features are provided by reading temperature sensors and other measuring instruments. Some proposed features, which called historical and control features, have been produced by preprocessing processes. Historical features have been created to keep the last changes in thermal parameters. Control features are used to add the possibility of predicting the thermal response generated by control decisions. Due to the high number of features, some of them are selected as thermal model inputs using two proposed methods with names square of correlation difference and extended SCD. The multi-laxyer perceptron neural network has been selected as the thermal model by comparing some regression models. The proposed thermal model has been compared with a number of important approaches to predict the temperature of 2 to 5 next seconds. The mean absolute error of the proposed model is calculated to be less than 0.5 and 0.7 ° C for the next 2 and 5 seconds, respectively. Furthermore, an online thermal model is proposed to increase the accuracy of prediction. In this approach, a thermal model is created in an offline step. Then, the model is updated at runtime for states where the model's accuracy is low. The proposed model consists of several thermal phases. For each phase, an MLP network is used to predict the temperature. Different thermal phases have been identified using the adaptive resonance theory network, according to the parameters affecting the processor temperature. Appropriate features are selected by the proposed algorithm for each thermal phase. The proposed model is capable of adding a new thermal phase to a set of phases and creating a suitable neural network for it at runtime. The proposed model error is less than 1 ° C for different time distances. Finally, using the proposed thermal model and a control model, the processor temperature is controlled by determining the processor frequency and fan speed. Proper features are selected for the controller. The error of the thermal control model in determining the processor frequency and fan speed is 2% and 0.6% respectively.