為了提高工具機熱變形位移模型之預測精與降低模型訓練時間，本文提出兩種自迴歸移動平均(auto regression moving average, ARMA)與前向式神經網路(feed-forward neural network, FNN)混成之預測模型。混成模型中ARMA模型被用來作溫度量測值與位移量測?的前處理，其輸出顧FNN之輸入，因此比統神經網路(neural network, NN)直接以溫度為輸入的方法，可大為減少網路節點數。有以傳統的ARMA方法，FNN方法與所提出的方法，以實機量測之舒測結果進行比較，探討本方法之預測精度與可行性。由預測結果得知，在相同之迭代次數下，兩種混成模型除了可降低模型之訓練時間外，並獲得較佳之預測精度。

A hybrid model consisted of ARMA (auto regression moving average) model and the FNN (feed-forward neural network) are proposed to increase the prediction accuracy and reduce the learning time for estimation of the thermal deformation in machine tools. The ARMA model is used to preprocess the measured temperatures and thermal deformations, and its outputs are treated as the inputs of the FNN. The proposed hybrid model can describe the relationships between the variations of measured temperatures and the thermal deformations. The experimental results show that the prediction accuracy of hybrid model is better than that of the conventional ARMA model or the FNN model with the same learning iterations.