Fault-tolerant voltage control framework for DC islanded microgrid: A synthesis of HS-ESN and time series analysis

• The proposed algorithm can instantly identify faulty sensors, unlike previous works that treat estimated values as healthy without diagnostic approaches. • Unlike existing methods that only detect sensor faults or cyber-attacks, this strategy both detects and alleviates sensor faults. • The approach does not rely on MPC/PI controllers, avoiding added complexity and the need for a default healthy sensor as a reference. • The proposed echo state network offers a simple, accurate, and computationally efficient structure that does not require historical data or fault training data, unlike other neural networks. Sensor fault detection and mitigation are critical for the safe, efficient, and stable operation of DC microgrids. This paper shows how the cooperation of artificial neural networks and time series analysis can help in detecting and mitigating voltage sensor faults in a DC islanded microgrid. Firstly, an echo state network (ESN) estimates the voltage value by the use of a current sensor embedded in the point of common coupling. Then, if the measured value differs from the estimated value, a time series-based identification algorithm recognizes the faulty sensor and applies the correct value to control the DC microgrid voltage. The proposed ESN estimator, which the harmony search optimization algorithm is used to select its parameters optimally, does not use historical inputs and provides significantly better performance compared to the nonlinear auto-regressive exogenous network, which has recently been the focus of numerous researchers. Comparison criteria include lower voltage ripples and fluctuations in steady-state conditions, lower maximum estimation error in a transient manner (0.17 % versus 3.41 %), and lower mean square error value in the testing stage of the network (6.34e-6 versus 1.04e-4). Moreover, the proposed faulty sensor identification algorithm discards the assumption that the estimated value is always correct. Consequently, it reduces the computational effort by eliminating the need for the additional proportional-integral controller. The conducted case studies confirm that the proposed framework effectively identifies all kinds of constant and time-varying sensor faults along with noise on the sensor, and consequently prevents erroneous values from being injected into the system.