<p>One of the challenges of the Internet of Things and smart cities is energy consumption and energy theft. An accurate approach to predicting energy consumption and detecting energy theft in smart cities increases efficiency and energy efficiency. Forecasting energy consumption makes energy production based on the needs of consumers, and detecting energy theft makes energy consumption forecasting models more accurate. In this manuscript, in the first step, the data set is balanced using the generative adversarial network based on game theory and the synthetic minority oversampling based on sample density method. In the second step, the basic features of the samples are selected with the Namib beetle optimization (NBO) algorithm to reduce the input of the CNN-LSTM model. In the third step, the hyperparameters of the CNN-LSTM model are optimized to reduce the prediction and classification error rate with the NBO algorithm. In the Benin Electricity Company dataset, the proposed method has fewer errors in predicting energy consumption than the LSTM, GRU, ARIMA-LSTM, and ARIMA-GRU methods. On the Individual Household Electric-Power Consumption dataset, the proposed method provides lower energy consumption prediction errors than convolutional neural network (CNN), long short-term memory (LSTM), and CNN-LSTM. The NBO algorithm optimizer CNN-LSTM hyperparameters more accurately than Coati optimization algorithm, jellyfish&#xa0;search optimization, Harris hawks optimization (HHO), and African vultures optimization algorithm. Experiments on the State Grid Corporation of China dataset showed that the proposed method's accuracy, sensitivity, and precision in predicting energy theft are 98.93, 98.32, and 96.78%. The proposed method is more accurate than CNN, DeepCNN, CNN-LSTM, and the gated recurrent unit (GRU) method.</p>