This paper presents a maximum power point tracking (MPPT) technique for high-performance wind generators with induction machines based on the growing neural gas (GNG) network. In this paper, a GNG network has been trained offline to learn the turbine-characteristic surface torque versus wind and machine speeds and has been implemented online to obtain the wind tangential speed on the basis of the estimated torque and measured machine speed (surface function inversion). The machine reference speed is then computed on the basis of the optimal tip speed ratio. For the experimental application, a back-to-back configuration with two voltage source converters has been considered, one on the machine side and the other on the grid side. The field-oriented control of the machine has been further integrated with an intelligent sensorless technique; in particular, the so-called total least squares (TLS) EXIN full-order observer has been adopted. Finally, a comparison with a classic perturb-and-observe MPPT has been made on a real wind-speed profile.

This paper presents a single-phase shunt active power filter (APF) for current harmonic compensation based on neural filtering. The shunt active filter, realized by a current-controlled inverter, has been used to compensate a nonlinear current load by receiving its reference from a neural adaptive notch filter. This is a recursive notch filter for the fundamental grid frequency (50 Hz) and is based on the use of a linear adaptive neuron (ADALINE). The filter's parameters are made adaptive with respect to the grid frequency fluctuations. A phase-locked loop system is used to extract the fundamental component from the coupling point voltage and to estimate the actual grid frequency. The current control of the inverter has been performed by a multiresonant controller. The estimated grid frequency is fed to the neural adaptive filter and to the multiresonant controller. In this way, the inverter creates a current equal in amplitude and opposite in sign to the load harmonic current, thus producing an almost sinusoidal grid current. An automatic tuning of the multiresonant controller is implemented, which recognizes the largest three harmonics of the load current to be compensated by the APF. The stability analysis of the proposed control system is shown. The methodology has been applied in numerical simulations and experimentally to a properly devised test setup, also in comparison with the classic sinusoidal current control based on the P-Q theory.

This paper deals with a single-phase distributed generation (DG) system with active power filtering (APF) capability, devised for utility current harmonic compensation. The idea is to integrate the DG unit functions with shunt APF capabilities, because the DG is connected in parallel to the grid. With the proposed approach, control of the DG unit is performed by injecting into the grid a current with the same phase and frequency of the grid voltage and with an amplitude depending on the power available from renewable sources. On the other hand, load harmonic current compensation is performed by injecting into the alternating current system harmonic currents like those of the load but with an opposite phase, thus keeping the line current almost sinusoidal. Both detection of the grid voltage fundamental and computation of the load harmonic compensation current have been performed by two neural adaptive filters with the same structure, one in a configuration ldquonotchrdquo and the other in the complementary configuration ldquoband.rdquo The ldquonotchrdquo filter has been used to compute the compensation current by eliminating only the contribution of the fundamental of the load current, whereas the ldquobandrdquo configuration is able to extract the fundamental of the coupling point voltage. Furthermore, because the active power generation and the APF features require current control of components at different frequencies, respectively, a multiresonant current controller has been adopted. The methodology has been tested successfully both in numerical simulation and experimentally on a suitably devised test setup. The stability analysis of the proposed control approach has been performed in the discrete domain.

This paper deals with a new experimental approach to the parameter estimation of induction motors with least-squares techniques. In particular, it exploits the robustness of total least-squares (TLS) techniques in noisy environments by using a new neuron, the TLS EXIN, which is easily implemented online. After showing that ordinary least-squares (OLS) algorithms, classically employed in the literature, are quite unreliable in the presence of noisy measurements, which is not the case for TLS, the TLS EXIN neuron is applied numerically and experimentally for retrieving the parameters of an induction motor by means of a test bench. Additionally, for the case of very noisy data, a refinement of the TLS estimation has been obtained by the application of a constrained optimization algorithm which explicitly takes into account the relationships among the K-parameters. The strength of this approach and the enhancement obtained is fully demonstrated first numerically and then verified experimentally.

This paper proposes a neural network (NN) model reference adaptive system (MRAS) speed observer suited for linear induction motor (LIM) drives. The voltage and current flux models of the LIM in the stationary reference frame, taking into consideration the end effects, have been first deduced. Then, the induced part equations have been discretized and rearranged so as to be represented by a linear NN (ADALINE). On this basis, the transport layer security EXIN neuron has been used to compute online, in recursive form, the machine linear speed. The proposed NN MRAS observer has been tested experimentally on suitably developed test set-up. Its performance has been further compared to the classic MRAS and the sliding-mode MRAS speed observers developed for the rotating machines.

This paper proposes a neural based full-order Luenberger adaptive speed observer for sensorless linear induction motor (LIM) drives, where the linear speed is estimated with the total least squares (TLS) EXIN neuron. A novel state space-vector representation of the LIM has been deduced, taking into consideration its dynamic end effects. The state equations of the LIM have been rearranged into a matrix form to be solved, in terms of the LIM linear speed, by any least squares technique. The TLS EXIN neuron has been used to compute online, in recursive form, the machine linear speed. A new gain matrix choice of the Luenberger observer, specifically taking into consideration the LIM dynamic end effects, has been proposed, overcoming the limits of the gain matrix choice based on the rotating-induction-machine model. The proposed TLS full-order Luenberger adaptive speed observer has been tested experimentally on an experimental rig. Results have been compared with those achievable with the TLS EXIN MRAS, the classic MRAS, and the sliding-mode MRAS observers.