Electricity theft is a severe non-technical losses for electricity provider, as it reduces supply quality, forces legitimate consumers to pay higher bills, and negatively impacts on the economy. The development of electricity theft detection models play a major role in identification of electricity theft consumers. The main issue with current anomaly detection models are the misclassification of imbalanced energy consumption data, high false positive rates and overfitting issues. An Intensive research is essential for the precise identification of electricity theft and for recovering significant economy losses faced by utility companies. To this end, a hybrid convolutional neural network- bidirectional long short-term memory (CNN-BiLSTM) with attention mechanism based deep learning model for anomaly detection is proposed in this work. The proposed model addresses the data imbalance problem by employing two data balancing approaches: K-means Synthetic Minority Over-sampling Technique (K-SMOTE) and K-means Adaptive Synthetic Sampling (K-ADASYN). The proposed model utilizes real time data I.e, State Grid Corporation of China (SGCC) to analyze and address electricity theft, and also it compares with different classification and deep learning based detection models such as Support vector machine (SVM), Logistic regression (LR), Long shortterm memory (LSTM), Bidirectional long short-term memory (BiLSTM), Convolutional neural network (CNN), and hybrid models such as CNN-LSTM, CNN- BiLSTM, CNN-LSTMAttention mechanism. To verify the proposed model different performance metrics are considered such as precision, recall, F1 score, accuracy, Matthews Correlation Coefficient (MCC), and Area Under Curve (AUC). From the results, it is clearly evident that the proposed CNN-BiLSTM-Attention mechanism with Kmeans ADASYN has better accuracy of 95%, F1 score of 95%, Recall of 95%, Precision of 95%, MCC of 90% and AUC of 95% When compared with other balancing approach and different algorithms.