Academia.eduAcademia.edu

Figure 2 – uploaded by Vanessa Becerra

See full PDFdownloadDownload figure
Fig. 3 — Schematic representation of the test station. (BP) learning method is used to minimize the least mean square (LMS) error function, e, is introduced, according to

Figure 3 — Schematic representation of the test station. (BP) learning method is used to minimize the least mean square (LMS) error function, e, is introduced, according to

Related Figures (25)

Fig. 1 — Multilayer perceptron network topology. The ANNs learning is a process by which the internal parameters (weights) of a neural network are continuously adjusted due the stimulation from the changing environment in which the network is immersed. The training task make possible to adjust speed and accuracy of learning the input—output relationships. The multilayer networks are universal tools capable of approximating any measurable function to any desired degree of accuracy. In this work a multilayer Perceptron network (MLP) was used as a platform to build the proposed models (see Fig. 1). A MLP consist in arrays of units divided in layers with only forward connections to units in subsequent layers. The connections have weights associated with them, therefore each signal traveling through the link is weighted. The first layer is the input layer, and the input units distribute the inputs to units in subsequent layers. The layers between the input layer and output layer are called hidden layers, the purpose of these hidden layers is to empower the information process, however there is no theory available to determine the correct amount of layers and units [7,8]. The type of learning is determined by the manner in which the parameter changes take place. The backpropagation This paper describe the development of a hybrid power plant model based on artificial intelligence, a combination of three renewable energy sources contributes to satisfy the energy demand for a remote housing application, minimizing costs and guaranteeing the correct system operation. The model considers the PV, WT and hydrogen-based systems operating in meteorological conditions in the state of Zacatecas
Fig. 1 — Multilayer perceptron network topology. The ANNs learning is a process by which the internal parameters (weights) of a neural network are continuously adjusted due the stimulation from the changing environment in which the network is immersed. The training task make possible to adjust speed and accuracy of learning the input—output relationships. The multilayer networks are universal tools capable of approximating any measurable function to any desired degree of accuracy. In this work a multilayer Perceptron network (MLP) was used as a platform to build the proposed models (see Fig. 1). A MLP consist in arrays of units divided in layers with only forward connections to units in subsequent layers. The connections have weights associated with them, therefore each signal traveling through the link is weighted. The first layer is the input layer, and the input units distribute the inputs to units in subsequent layers. The layers between the input layer and output layer are called hidden layers, the purpose of these hidden layers is to empower the information process, however there is no theory available to determine the correct amount of layers and units [7,8]. The type of learning is determined by the manner in which the parameter changes take place. The backpropagation This paper describe the development of a hybrid power plant model based on artificial intelligence, a combination of three renewable energy sources contributes to satisfy the energy demand for a remote housing application, minimizing costs and guaranteeing the correct system operation. The model considers the PV, WT and hydrogen-based systems operating in meteorological conditions in the state of Zacatecas
Fig. 2 — Adapting FIR-MLP schema. (a) Signal-flow graph of a connection FIR unit; (b) open loop (training) and (c) close loop (recursive prediction).
Fig. 2 — Adapting FIR-MLP schema. (a) Signal-flow graph of a connection FIR unit; (b) open loop (training) and (c) close loop (recursive prediction).
Table 1 — PEM-FC input and output variables and ranges after discrimination and separation accordance. The weights, which at the beginning are random numbers close to zero, are successively updated in the direction of the decreasing error gradient. The major assumption is that the weight correction is proportional to this gradient, with a constant, also known as the learning rate: the learning process is one way to achieve it. A popular technique used is called TDNN (Time Delay Neural Network) [10]. The TDNN is an ANN with forward propagation whose hidden and output neurons are replicated across time. The MLP consider a network whose layers of hidden neurons and output are all based on the impulse response of finite duration FIR (finite-duration impulse response), this network is called a Multilayer Perceptron Network FIR (FIR-MLP) [11]. Fig. 2 a shows the FIR connection schema, where wij (1) denote the weight connected to the Ith tap of the FIR filter modeling the connection between the neuron i and the neuron j; the index 1 ranges from 0 to M (number of delay units); the signal x;(n) appearing at the output is given by a linear combination of delayed values of the input signal x;(n).
Table 1 — PEM-FC input and output variables and ranges after discrimination and separation accordance. The weights, which at the beginning are random numbers close to zero, are successively updated in the direction of the decreasing error gradient. The major assumption is that the weight correction is proportional to this gradient, with a constant, also known as the learning rate: the learning process is one way to achieve it. A popular technique used is called TDNN (Time Delay Neural Network) [10]. The TDNN is an ANN with forward propagation whose hidden and output neurons are replicated across time. The MLP consider a network whose layers of hidden neurons and output are all based on the impulse response of finite duration FIR (finite-duration impulse response), this network is called a Multilayer Perceptron Network FIR (FIR-MLP) [11]. Fig. 2 a shows the FIR connection schema, where wij (1) denote the weight connected to the Ith tap of the FIR filter modeling the connection between the neuron i and the neuron j; the index 1 ranges from 0 to M (number of delay units); the signal x;(n) appearing at the output is given by a linear combination of delayed values of the input signal x;(n).
Fig. 4 — Multilayer perceptron based on the finite impulse response. (left) Dynamic neural model of the PEM-FC and FIR-MLP correlation rates recursive prediction of the stack voltage variable (left) and temperature (right) performed by the neural model and compared against the real PEM-FC system. To induce dynamic properties to the model, memory capacity must be provided to the ANN [9], a set of time delays units in the proposed topology and their values adjustment in A fuel cell (FC) is an electro-chemical device which converts the chemical energy of hydrogen or hydrogen containing fuel, such as a hydrocarbon, alcohol or other, to electricity by means of an electrochemical reaction [12]. Electrically an FC behaves as a current source whose output voltage varies depending of the current drawn by the external load. The
Fig. 4 — Multilayer perceptron based on the finite impulse response. (left) Dynamic neural model of the PEM-FC and FIR-MLP correlation rates recursive prediction of the stack voltage variable (left) and temperature (right) performed by the neural model and compared against the real PEM-FC system. To induce dynamic properties to the model, memory capacity must be provided to the ANN [9], a set of time delays units in the proposed topology and their values adjustment in A fuel cell (FC) is an electro-chemical device which converts the chemical energy of hydrogen or hydrogen containing fuel, such as a hydrocarbon, alcohol or other, to electricity by means of an electrochemical reaction [12]. Electrically an FC behaves as a current source whose output voltage varies depending of the current drawn by the external load. The
Table 2 — Best linear fit and calculated errors in the validation process of the dynamic neural model.
Table 2 — Best linear fit and calculated errors in the validation process of the dynamic neural model.
Fig. 5 — Schematic representation of the house made test station and electrolyzer main set of variables and their distribution in its operation space.
Fig. 5 — Schematic representation of the house made test station and electrolyzer main set of variables and their distribution in its operation space.
Table 3 — Best linear fit, linear deviation and correlation rates in the validation process for every cell of the PEM- WE stack.
Table 3 — Best linear fit, linear deviation and correlation rates in the validation process for every cell of the PEM- WE stack.
Fig. 6 — Dynamic model of a PEM-WE based in a FIR-MLP and correlation rates obtained from the validation sequence by cell.
Fig. 6 — Dynamic model of a PEM-WE based in a FIR-MLP and correlation rates obtained from the validation sequence by cell.
Fig. 8 — Time Delay Neural Network training diagram. The correlation values of the MHW and GW are very close, although MHW is slightly better and it is chosen to build the
Fig. 8 — Time Delay Neural Network training diagram. The correlation values of the MHW and GW are very close, although MHW is slightly better and it is chosen to build the
Fig. 7 — Signal reconstruction by MHW and GW for one single day solar radiation.
Fig. 7 — Signal reconstruction by MHW and GW for one single day solar radiation.
Table 5 — Best linear fit and linear deviation for the thre network processes.
Table 5 — Best linear fit and linear deviation for the thre network processes.
Fig. 9 — Predicted GSR days from de neural model vs real measured GSR days (left); annual prediction of the neural model during 2007. Wind turbines (WT), perform the conversion of wind kinetic energy into electrical energy. The WT classification is related to the power generation capacity, considering the nominal power. Small turbines are subdivided in micro-turbines (0-5 kW) and mini-turbines (5-80 kW). This work uses a micro-WT prototype of 1.5 kW (Fig. 10) developed by Tozzi Nord (technical specifications are shown in Table 7). The WT
Fig. 9 — Predicted GSR days from de neural model vs real measured GSR days (left); annual prediction of the neural model during 2007. Wind turbines (WT), perform the conversion of wind kinetic energy into electrical energy. The WT classification is related to the power generation capacity, considering the nominal power. Small turbines are subdivided in micro-turbines (0-5 kW) and mini-turbines (5-80 kW). This work uses a micro-WT prototype of 1.5 kW (Fig. 10) developed by Tozzi Nord (technical specifications are shown in Table 7). The WT
Table 6 — MSE calculated values for every month of the year reserved to validate the model. fault tolerance, generalization and contextual response prop- erties of ANNs allows building a model scalable and adaptable to similar weather conditions in Zacatecas. Training of measurements, which includes wind speed (V), direction (6) and generated power pdf(V) (see Fig. 11), were obtained from the three mentioned weather stations: station 1 located at the top of a 15 m building at the north of Rome (acquisition period: November 29th 2009—March 2nd 2010); second station located at Trento on a building at open field (acquisition period: November 24th—march 7th 2010); and the third one in on a building in Messina at 5.5 m from the floor (November 10th 2009—December 15th 2010). The dataset from station 1 and 2 are designated to train the MLP, and the third one is reserved to validate the model. The MLP neural model is built in three layers, activation function was determinate experimentally and comparing results (Fig. 12 left).
Table 6 — MSE calculated values for every month of the year reserved to validate the model. fault tolerance, generalization and contextual response prop- erties of ANNs allows building a model scalable and adaptable to similar weather conditions in Zacatecas. Training of measurements, which includes wind speed (V), direction (6) and generated power pdf(V) (see Fig. 11), were obtained from the three mentioned weather stations: station 1 located at the top of a 15 m building at the north of Rome (acquisition period: November 29th 2009—March 2nd 2010); second station located at Trento on a building at open field (acquisition period: November 24th—march 7th 2010); and the third one in on a building in Messina at 5.5 m from the floor (November 10th 2009—December 15th 2010). The dataset from station 1 and 2 are designated to train the MLP, and the third one is reserved to validate the model. The MLP neural model is built in three layers, activation function was determinate experimentally and comparing results (Fig. 12 left).
power availability function that the model describes. From the obtained results, and comparing similar works in literature [45,46], this model presents good prediction accuracy on wind energy availability in the range of 4-15 m s-*. However, in the zone of 15-20 ms * a considerable deviation is observed due to the unavailability of experimental data. Despite this limi- tation, and according to the measurements obtained from the site of interest, this model can be used in the entire hybrid plant model.
power availability function that the model describes. From the obtained results, and comparing similar works in literature [45,46], this model presents good prediction accuracy on wind energy availability in the range of 4-15 m s-*. However, in the zone of 15-20 ms * a considerable deviation is observed due to the unavailability of experimental data. Despite this limi- tation, and according to the measurements obtained from the site of interest, this model can be used in the entire hybrid plant model.
Fig. 10 — Vertical axis micro-wind turbine of 1.5 kW (left) and its weather station installed (right) A linear regression (see Fig. 13 left) was calculated in order to verify the correlation of the neural model and the real system, obtaining a rate of R = 0.88 and MSE = 14%. The energy available was calculated and compared against the Weibull law function given by the manufacturer Tozzi Nord Inc. Fig. 13 (right) shows the similarity between the energy lines of the model and of the manufacturer, however, since wind speed values above 14m s * were not measured in none of the three weather stations, the MLP model was not able to follow the Last sections describe the developed models that integrate the hybrid plant. This work establishes a new methodology to design and size a hybrid power plant based on a PV system, a Wind micro-Turbine, a PEM-FC system and a PEM-WE for
Fig. 10 — Vertical axis micro-wind turbine of 1.5 kW (left) and its weather station installed (right) A linear regression (see Fig. 13 left) was calculated in order to verify the correlation of the neural model and the real system, obtaining a rate of R = 0.88 and MSE = 14%. The energy available was calculated and compared against the Weibull law function given by the manufacturer Tozzi Nord Inc. Fig. 13 (right) shows the similarity between the energy lines of the model and of the manufacturer, however, since wind speed values above 14m s * were not measured in none of the three weather stations, the MLP model was not able to follow the Last sections describe the developed models that integrate the hybrid plant. This work establishes a new methodology to design and size a hybrid power plant based on a PV system, a Wind micro-Turbine, a PEM-FC system and a PEM-WE for
Fig. 11 — Statistic information provided by the weather station: distribution of frequency density and wind direction distribution (left) and aero-generator TN 1.5 delivery power behavior (right).
Fig. 11 — Statistic information provided by the weather station: distribution of frequency density and wind direction distribution (left) and aero-generator TN 1.5 delivery power behavior (right).
Fig. 12 — MLP based model of the micro-WT and its predicted power supply under real meteorological conditions
Fig. 12 — MLP based model of the micro-WT and its predicted power supply under real meteorological conditions
Fig. 13 — Left: Correlation rate calculated from the neural model prediction, linear deviation Am = 0.022 y Ab = 2.259; Right: Wind power distribution of the aero-generator TN 1.5, comparison of the manufacturer’s Weibull function against the power distribution prediction of the MLP model. 2.6.1. System description and methodology A dynamic model of the hybrid power plant is illustrated in the following schema (Fig. 14 left), although it considers In order to supply the power demand and to overcome the energy storage problem, hydrogen can be produced and storage for mid-term use. The PEM-WE uses the generated power from the PV and WT systems to produce hydrogen,
Fig. 13 — Left: Correlation rate calculated from the neural model prediction, linear deviation Am = 0.022 y Ab = 2.259; Right: Wind power distribution of the aero-generator TN 1.5, comparison of the manufacturer’s Weibull function against the power distribution prediction of the MLP model. 2.6.1. System description and methodology A dynamic model of the hybrid power plant is illustrated in the following schema (Fig. 14 left), although it considers In order to supply the power demand and to overcome the energy storage problem, hydrogen can be produced and storage for mid-term use. The PEM-WE uses the generated power from the PV and WT systems to produce hydrogen,
Table 8 — Electric specifications of PV module at 25 °C and solar radiation of 1000 wm ~?. then, this fuel acts as an energy buffer when is employed by the PEM-FC system. In that sense the power supply fluctuation from the PV and wind systems are tackled by this media.
Table 8 — Electric specifications of PV module at 25 °C and solar radiation of 1000 wm ~?. then, this fuel acts as an energy buffer when is employed by the PEM-FC system. In that sense the power supply fluctuation from the PV and wind systems are tackled by this media.
Fig. 14 — Hybrid power plant schema (left) and a fuzzy logic controller for power management strategy (right).
Fig. 14 — Hybrid power plant schema (left) and a fuzzy logic controller for power management strategy (right).
Fig. 15 — Current—voltage curves of the PV system for different values of short circuit current. In the integral power plant, measurements and status information of each component and its corresponding DC/DC converter are sending to the controller, which evaluates them and delivers instructions for the energy management strategy. PV arrays considers information provided by the manu- facturer (see Table 8) and it is used to calculate I—V curves and the Maximum Power Point (MPP) which varies according to solar radiation and current demand (Fig. 15). According to the power curves of the PV system and the GSR predictive model
Fig. 15 — Current—voltage curves of the PV system for different values of short circuit current. In the integral power plant, measurements and status information of each component and its corresponding DC/DC converter are sending to the controller, which evaluates them and delivers instructions for the energy management strategy. PV arrays considers information provided by the manu- facturer (see Table 8) and it is used to calculate I—V curves and the Maximum Power Point (MPP) which varies according to solar radiation and current demand (Fig. 15). According to the power curves of the PV system and the GSR predictive model
Table 9 — Main electrical devices considered in the remot housing application. described before, PV system performance can be simulated under weather conditions in the interested site.
Table 9 — Main electrical devices considered in the remot housing application. described before, PV system performance can be simulated under weather conditions in the interested site.
A fuzzy logic controller relates an input—output space of variables trough a list of if-then rules. Within these conditional sentences, input concepts (fuzzy sets) are defined as membership functions. The if-part (called antecedent) specifies an operation condition (input combination). The then-part of a rule (consequent) refers to a control action. This control command is defined by the degrees of membership of the if- part, therefore the then-part of all rules are averaged and weighted by these membership degrees. Fig. 16 — Performance of the hybrid power plant for a typical day in Zacatecas: (a) PV power supply; (b) Wind turbine power supply; (c) PEM-FC power supply; (d) hydrogen production and (e) random load profile and hybrid plant general performance. Since the PV-WT systems are the main source of power, the starting point is defined by
A fuzzy logic controller relates an input—output space of variables trough a list of if-then rules. Within these conditional sentences, input concepts (fuzzy sets) are defined as membership functions. The if-part (called antecedent) specifies an operation condition (input combination). The then-part of a rule (consequent) refers to a control action. This control command is defined by the degrees of membership of the if- part, therefore the then-part of all rules are averaged and weighted by these membership degrees. Fig. 16 — Performance of the hybrid power plant for a typical day in Zacatecas: (a) PV power supply; (b) Wind turbine power supply; (c) PEM-FC power supply; (d) hydrogen production and (e) random load profile and hybrid plant general performance. Since the PV-WT systems are the main source of power, the starting point is defined by
Related topics:
EngineeringEnvironmental ScienceComputer ScienceRenewable EnergyPhotovoltaic system
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved