Carbonate fracture-cavity reservoirs are characterized by strong heterogeneity. Single seismic attribute prediction and conventional seismic fusion methods do not take additional information such as mud leakage into consideration，which could lead to large errors. According to well seismic calibration，a fracture-cavity reservoir prediction method based on well-controlled multi-attribute machine learning was proposed to finely predict fracture-cavity reservoirs，with the attributes of mud-leakage points taken as the constraints. Firstly，in accordance with the results of well seismic calibration，different sensitive seismic attribute values at the leakage points were extracted as input data，and the reservoir types defined by the features of leakage points were output to form training set data. Then，based on the support vector machine（SVM）method，model training was conducted on the data to obtain a prediction model highly consistent with prior seismic attributes. Finally，the model was applied to predict the Ordovician fracture-cavity reservoir in Shunbei Oilfield，Tarim Basin.The prediction results show that this method can reflect the real reservoir characteristics and fits well with drilling features.