Wind energy has become a key driver of economic growth in Spain, as evidenced by the growing number of wind farms, with approximately 22,000 wind turbines spread across 1,500 installations. Despite its benefits, the proliferation of wind farms poses a substantial threat to bat populations as it is estimated that, in Spain alone, one million bats per year may fall victim to the blades of these turbines. Current methods of monitoring bat populations involve laborious analysis of huge ultrasonic data sets, which poses a significant challenge due to the diversity of bat species, background noise, and the different characteristics of their calls. To address this problem, this work proposes the development of deep learning models capable of accurately discerning bat calls from ambient noise. More than 150 models were evaluated, incorporating various neural network architectures and combinations of hyper-parameters. The five best models were selected to form an ensemble, called BatNoiseDL, which achieved an accuracy rate of \(97.1\%\) on 500 difficult audio files where bat identification was especially challenging due to high noise levels. In addition, BatNoiseDL doubled the prediction improvement over three popular commercial tools on an audio sample of 5618 where these tools failed in a high percentage of cases. Therefore, comparative analysis with existing software tools highlighted the superior filtering capability and efficiency of the proposed method against them. The potential of this method is not only accuracy, but also speed improvement through the use of dedicated GPU servers.