Artificial Intelligence promises to revolutionize the prognosis of complex systems like gas turbines and facilitate their long-term autonomous operation by enabling both predictive maintenance and recovery from potential faults via a suitable re-configuration of the deployed system. We present a recurrent neural network based approach for probabilistic prediction of the power generated from a gas turbine up to 5 minutes in the future by observing (i) the generated power in the past and (ii) past snapshots of 80 other sensor data, including those monitoring vibrations. Our deep neural network with 367,041 parameters is trained on a system containing 2 NVIDIA GPUs RTX 2080 with 5,888 CUDA cores, 32 CPU cores, and 128 GB RAM, and our neural network achieves a root mean square error of 0.0029 in 10 epochs in about 14.5 minutes. The deep neural network produces a root mean square error of 0.0028 on a validation data set. We also define an event of interest to be a drop of the generated power from 70% of the normal to a value below 50% of the normal value. Our test data set contains 6 such events of interest. Our recurrent neural network is trained with the objective of making predictions up to 5 minutes in the future; we observed that the neural network can predict all of these six events of interest in the test data before their occurrence.
