Cloud workload data is time series data that consists of the logs of the resource consumption, such as central processing unit (CPU) utilization, memory utilization and bandwidth consumption, on a particular machine in a cloud environment. These logs are essential for analyzing the machine’s performance and debugging purposes at fault time. It is worth noting that the utilization fluctuates due to changes in load over time. Moreover, there are some peaks and valleys in the utilization pattern, and predicting such points in the cloud workload is challenging and not well-studied. The prediction can be helpful in saving the waste of resources and providing quality services to the customers. Recently, researchers have used deep learning algorithms to predict the turning points using basic and fluctuating features. However, the prediction can be improved by taking the extreme points of the adjacent segments, and it is effective for high-variance time series data. This paper presents a turning point prediction (TPP) algorithm on cloud workload data using deep learning models, namely long short-term memory (LSTM), bidirectional LSTM (BiLSTM) and gated recurrent unit (GRU) networks, that use some novel features. TPP uses a piece-wise linear segmentation (PLS) algorithm to generate the segments. Then, it determines the potential turning points and labels them into two classes. Finally, the LSTM, BiLSTM, and GRU are applied to classify the turning points. We simulated the TPP on two datasets, Google cloud and Alibaba, and compared its three variants in terms of precision, recall and F1-score to show its effectiveness.