In wireless mesh networks (WMNs), traffic prediction refers to predicting traffic volumes, which comprises transmitting data and receiving requests by network nodes in WMN. However, existing prediction methods have failed to predict network traffic accurately due to the dynamic nature and complex pattern of WMN. To overcome this problem, a multi-head attention with long short-term memory (MHA-LSTM) is employed for a prediction model in network prediction for wireless mesh networks. The dataset used for prediction was collected from eight sensors, which are placed in high-speed diesel (HSD) pump. Initially, sensor data are fed to preprocessing to check stationary of time series data by ADFuller and fed to the proposed MHA-LSTM prediction model, which explores high-level temporal features of WMN, and the predicted output will be determined by the SoftMax layer of LSTM. The experimental analysis indicates promising results for network prediction with mean average error (MAE) of 9.12%, root mean-squared error (RMSE) of 13.35% and mean average prediction error (MAPE) of 7.12% for 60 min, which is higher than existing prediction methods like convolutional-gated recurrent unit (Conv-GRU), LSTM, and bidirectional-LSTM (Bi-LSTM).