A hybrid grid (HG) is an electrical distribution network that allows the penetration of various locally produced sources with or without storage equipment (Navigant Research, 2016). Reliability and cost savings can be achieved by the use of renewable energy sources (RESs), traditional turbines, storage plants, and electricity charges, as well as the use of an HG. A hybrid grid can work in both connected and isolated modes. The HG can be linked to the main grid in the grid-connected mode via a common interconnector point and participate in energy trading as either a customer or a provider via grid receipt or power transmission.

A power management system (PMS) is also used for the most efficient use of hybrid grid units. The PMS techniques can maximize the transmission of power generation outputs from HG units and ensure economic load demand (LD), as well as monitor the frequency and voltage of HG systems. Rauf et al. (2016) outlined the latest HG-PMS architecture and the different power and heat generation systems and electric cars, as well as the core functionality and limitations of HG-PMS and the use of optimization technology. Photovoltaic (PV) and wind power are the most commonly used sources of renewable energy in HGs. However, because of the intermittent nature of renewable energy sources, it is recommended that it be assisted by appropriate storage units and optimally integrated into the HG scheme. In terms of the maximum use of the battery storage (BS) in the HG, several studies were carried out for optimal delivery of HG generation and storage devices. For this reason, many advanced methods of optimization were published. A combined hybrid supply of energy and a battery source multimedia problem is solved with multi-target integral linear programming (Mathieu et al., 2013). Fuzzy logic control (FLC) determines the charging and discharge of the battery. Capital and maintenance expenditures, fuel prices, electricity costs, and carbon sanctions are the target functions to be minimized. The PMS solution for GC-HG is combined with the energy storage system (ESS) and local generation (PS) in a genetic algorithm (Rauf et al., 2016). The proposed PMS would determine how much HG electricity will be added to the ESS and how much will be replaced on the main grid. The primary aim is to maximize profits from power grid trade. Kasis et al. (2017a) investigated the planning of battery and HG generation plants. The system is modeled on a stochastic framework, and the PMS is solved using a deep Q-learning algorithm. In the study by Kasis et al. (2017b), we examined a dynamic dispatch in the HG of BS. The objective function of the problem is designed to maximize the battery operational benefit. To solve the battery management issue, a reinforcement study is proposed in combination with a search of Monte-Carlo trees. In the study by Schuitema et al. (2017), FLC was proposed to control the BS level of the state of charge (SOC) limited to the minimum and maximum limits. Moreover, a backtracking search algorithm is introduced to determine an optimum battery SOC control to allow for smooth BS operation throughout the time and to minimize the overloading and unchanging capacity. In the study by Books and Barooah (2019), an optimum power send for the GCHG was introduced with PV units and battery deactivators in the programming algorithm. The proposed goal is to maximize the amount of PV, minimize operating costs, and trade electricity costs with the grid. In the study by Kim et al. (2016), for a day-long plan, an optimal PMS solution was tested. The problem of optimization is modeling so that the operating costs for HG generation units are minimized and the consumption of renewable sources maximized. The optimization problem is fixed by convection programming, and a rolling horizon-predictive controller in combination with a predictive model determines the best environment for a BS. In the study by Aduda et al. (2018), the PMS of an isolated HG was targeted using a cuckoo-search algorithm with RES, DG, and BS. The target problem feature is to decrease overall costs and reduce gas emissions. Applying a GA achieves an economic advantage by using RES (Lin et al., 2015). The device studied is an isolated HG, BS, and DG with renewable sources. The goal feature proposed is to minimize capital and to maintain HG units, fuel costs, and pollution costs. In the GC-HG floral pollination algorithm (Chen et al., 2017), integrated RES, micro turbines, fuel cells, and BS are examined for the solution of PMS. The aim is to minimize BS operating costs, energy produced costs, the energy expenses of the grid exchanged, and costs for the demand response. In the study by Vrettos et al. (2016), an amended algorithm to check the ideal operating condition of BS in a population HG was proposed in a particle swarm optimization (PSO). An effective battery charging and discharge feature is added. Most of these research works do not take into account the objective purpose of reducing battery life. As shown, the loading and unloading processes have a major influence on BS lifetime. This article includes one of the objective functions that should be reduced, which is lifecycle deprivation costs. The genetic algorithm is used for charging and discharging cycles and quantifying the cost of battery deprivation (BDC). In order to resolve the unit engagement of a grid-connected HG, the effectiveness of the artificial neural network with genetic algorithms has been demonstrated. This study uses the ANN-GA algorithm for solving a problem of 48-h programming for isolated rural HG units. The following is an overview of the article. HG Modeling discusses modeling of HG devices; PMS Problem Formulation deals with proposed PMS, including potential objective functions and devices; ANN With Genetic Application deals with application of the ANN and genetic algorithms in the PMS; Discussion on Simulation Outcomes deals with the simulation and discussion; and Conclusion ends the article.

The hybrid grid (HG) includes RES, wind turbine (WT), BS, and DGs with PV panels.

The forward propagation in each ANN-layer is expressed as follows (Kang et al., 2012), (Lu et al., 2010) and (Eftekhari, 2017): • X j l = b j l + ∑ i = 1 N conv 1 D ( w i j l − 1 , S i l − 1 ) , (30) where X j l is the input, b j l is the bias of the j ^th neuron in layer l, S i l − 1 is the output of the neuron in layer l-1, w i j l − 1 is the kernel from the i ^th neuron in layer (l-1) to the j ^th neuron in layer l, and conv1D (….) is used for performing 1D convolution without zero-padding. The dimension of the input array, X j l , is less than the output array’s dimension, S i l − 1 . The output, Y j l , in layer l may be expressed by passing input X j l through activation function f (.), as follows: • Y j l = f ( X j l ) , (31) • S j l = Y j l ↓ ss , (32) where S j l represents the output of the j ^th neuron in layer l and ↓ ss stands for the down-sampling operation with a factor, ss.

The GA algorithm propagates the error from the output of the MLP-layer. Let l = 1 for the input layer and l = L for the output layer. Assume that Q _L is the number of classes in the output, for a given input vector; let its target be [t ₁ ,….., t Q L ] and output vectors, [ y 1 L , ….., y Q L L ]. In the output layer, the mean-squared error (MSE), E, is expressed as follows: E = ∑ i = 1 Q L   ( y i L − t i ) 2 . (33)

The delta error, Δ j l = ∂ E ∂ X j l , should be computed to calculate the derivative of E with respect to each parameter in the network. For updating all weights of neurons and the bias of that neuron in the preceding layer, the chain-rule of the derivative is used as follows: ∂ E ∂ w i j l − 1 = Δ j l   y i l − 1 , (34) ∘ ∂ E ∂ b j l = Δ j l . (35)

Therefore, the regular back propagation from the first MLP layer to the last ANN layer is performed as follows: ∂ E ∂ S j l = Δ S j l = ∑ i = 1 M ∂ E ∂ X i l + 1   ∂ X i l + 1 ∂ S j l = ∑ i = 1 M Δ i l + 1 w j i l . (36)

After performing the first back propagation from layer (l+1) to current layer l, the GA may carry to the input delta of ANN layer l, Δ j l . Assume that the zero-order up-sampled map is us j l = up ( s j 1 ) ; then the delta error is expressed as follows: Δ j l = ∂ E ∂ y j l ∂ y j l ∂ X j l = ∂ E ∂ u S j l ∂ u S j l ∂ y j l f ′ ( X j l ) = up ( Δ S j l ) β f ′ ( X j l ) , (37) where β = ( ss ) − 1 .

The GA of delta error ( Δ S j l Σ ← Δ i l + 1 ) is expressed as follows: Δ S j l = ∑ i = 1 M conv 1 Dz ( Δ i l + 1 , rev ( w ji l ) ) , (38) where rev (.) is used for array reversing and conv1Dz (…) is used for full 1D convolution performing with zero-padding.

The derivative of the error with respect to weight and bias may be expressed as follows: ∂ E ∂ w ij l = conv 1 D ( S j l ,   Δ i l + 1 ) , (39) ∂ E ∂ b j l = ∑ n Δ j l ( n ) . (40)

The weights and biases can be updated with learning rate ∝ using the following equations: w ij l − 1 ( t + 1 ) = w ij l − 1 ( t ) − ∝ ∂ E ∂ w ij l − 1 , (41) b j l ( t + 1 ) = b j l ( t ) − ∝ ∂ E ∂ b j l . (42)

This article expresses a 24-h horizon energy transmission cycle of an isolated HG with battery storage. In addition to the fuel cost, the objective feature proposes battery depletion costs. The question of energy transmission was to reduce the FC of DG and BDC to a minimum. In the proposed target function, a WF is implemented. The ANN-GA addresses the energy supply issue in order to ensure the best supply of isolated HGs such as the WT engine, PV, two DGs, and batteries. The results show that it is necessary to select the appropriate value of WF for the best cost function. WF also affects the state of the SOC battery. The results show that the SOC battery with WF equal to 250 remains more robust and the charging cycles decreased. The ED problem of HG generating units is performed on e = 250 in order to resolve the trade-off between the BDC and FC. The ANN-GA adds all HG units together to satisfy the LD