During startup, hydroelectric turbine blades undergo important deformations, which significantly reduce the device’s Remaining Useful Life (RUL ) through fatigue. Directly measuring these deformations can only be done for a dozen of trials during the commissioning of the turbines. Our objective during these trials is to find the optimal startup controller parameters which minimize the deformation peak-to-peak range while reaching the synchronous rotation speed in given time. After gathering some initial startups, we train a neural model to predict the deformation envelope from the turbine’s startup trajectory in the 2-D space defined by the rotational speed and the wicket gates opening. This model, coupled with a turbine dynamics surrogate simulator, defines a black box. Given a set of startup parameters, it outputs a cost based on the deformation amplitude, the model’s uncertainty, and the time to reach the synchronous speed. The NOMAD optimizer is then used to find the optimal parameters. First, in an active learning phase, the cost considers the model’s uncertainty so that the optimized parameters, when tested on the turbine, provide relevant data to improve the deformation model. Then, we only consider the deformation amplitude and time to synchronous speed to output the optimal startup parameters.

Maintenance of electrical transmission equipment is key to ensure the reliability of power supply. Maintenance tasks may be triggered, for example, by equipment failures (reactive maintenance), observed anomalies expected to lead to failures (condition-based maintenance), and time (periodic preventive maintenance). Periodic maintenance tasks are typically scheduled at fixed time intervals to detect ongoing degradation mechanisms and apply corrective measures as needed. Usually, engineers choose the period between tasks using their knowledge of failure mechanisms. In the context of electric transmission networks, the period between tasks should be a compromise driven by factors that include equipment reliability, maintenance costs (inspections, repairs, replacements...), the value of lost load (VoLL) and other risks inherent to power transmission (environment, health and safety...). Herein, an asset behaviour model, an event stochastic simulator, a power-flow simulator, and a risk model with a VoLL estimator are combined to quantify the total cost of periodic maintenance strategies. Then, a blackbox optimization solver searches for periodic maintenance strategies that minimize costs within specified constraints. Given that the event simulator uses a Monte Carlo method to output network states where equipment fails following set statistical distributions, the blackbox is non-deterministic. However, timely and meaningful results are obtained through proper adjustment of the number of Monte Carlo cycles, the length of the timespan simulated, and other parameters. This opens the way to multifidelity optimization, where these parameters are automatically adjusted during optimization. Ultimately, engineers may use this approach to select optimal periodic maintenance schedules that minimize the global risk for the network operator. The procedure is implemented with NOMAD, an open-source blackbox optimizer.

Hyperparameter optimization remains one of the main challenges in machine learning and often relies on manual tuning of hyperparameters, or the use of naive approaches such as grid search and random search. Unfortunately, manual hyperparameter tuning is a lengthy and inefficient process that heavily depends on the user’s expertise. Grid search and random search, on the other hand, are techniques that are not very efficient and require a significant amount of computation.

With an electric short-term demand forecasting model based on a Temporal Fusion Transformer (TFT) architecture, we studied the impact of using different optimization methods, including Bayesian Optimization and Hyperband, Latin hypercube, and Mesh Adaptive Direct Search (MADS), on the performance of the model. Results show that the MADS optimization method was the most efficient and that the choice of the objective function and constraints has a significant impact on the resulting optimized forecasting model.