Transfer learning algorithms are capable to apply previously learned knowledge in source domain, which alleviates much expensive efforts of knowledge recollection in target domain. But the knowledge in source domain is always imperfect due to redundant or contaminated information. To solve this problem, an ensemble filter-transfer learning (EFTL) algorithm based on the source knowledge reconstruction is proposed in this paper. First, a knowledge partition strategy based on model is developed to classify the source knowledge into different types. Then, the positive knowledge can be identified, which contributes to target domain with a rejection of the negative transfer. Second, a knowledge filter algorithm is introduced to filter out the redundant information in non-positive knowledge. Then, the non-positive knowledge can be reconstructed by this algorithm to prevent the loss of available information. Third, an ensemble transfer mechanism is established to realize the synchronous transfer of omnidirectional knowledge for the target domain. Finally, comparative experiments on model prediction in practical applications are provided to illustrate the dependability of EFTL.