Power load forecasting using an artificial neural network (ANN) that allows for fast convergence with precise forecasts is the focus of this study. To ascertain how much power will be required at a specific moment to supply end customers and how that demand will impact the utility grid, accurate load forecasting is crucial. By using the power load forecast, waste and inefficiency may be prevented and sufficient power can be made available to fulfill consumption demands.

As explained by Neelakanta and De Groff in [1] , an artificial neural network (ANN) is a mathematical model that has been developed as a computational tool based on the image of the biological brain complex. The Hessian matrix of the pertinent input data serves as the foundation for supervised training. Details provided by the Mazoon Electrical Company in Oman provide the pertinent data for this investigation. For the years 2007 and 2008, monthly load data is gathered for a specific Oman region known as Al Batinah. In other words, twenty-four months’ worth of temperature and humidity data—two whole years—must be entered into the ANN during training. The maximum, minimum, and average temperatures, as well as the maximum, minimum, and average humidity, are the six inputs. Megawatts of load data are the outputs. In other words, this load data serves as the training’s teacher. The intended output is comparable to that of a supervisory teacher.

The mathematical method used by the artificial neural network is one that converts data from an input space to an output space. In order to minimize the error—that is, the difference between the network’s actual output vector and the intended output vector—supervised training aims to iteratively update a set of connectivity weights that are introduced between the neural layers in the ANN.

A multi-layered feed-forward perceptron (MLFP) consisting of an input layer, one hidden layer, and one output is the test ANN used in the simulations ( Figure 1 ). Through linked inner and hidden layers, the input values advance. In order to squash the total output to a limited level, it is fed into a nonlinear sigmoidal function. The sigmoid-compressed output is then compared with a teacher value, which stands for the intended output aim.

A backpropagation gradient technique computes and applies the resulting error to the interconnection weights. In other words, an output, O_i, is indicated by the sigmoid-compressed value and compared to a teacher/supervisory (reference) value, T_i, which represents the intended output aim. After that, the error corresponding to (O_i − T_i) is backpropagated. This error is stated in terms of an error function, ε, which represents the mean-squared value of (O_i − T_i). When the error function is applied to the inter-connection weights, W_ij, the backpropagation (BP) technique, usually allows a gradient based on steepest descent, which alters the weight vector values (either increasing or decreasing).

The rth ensemble of inputs {y_i}, weighted across the input-layer (i = 1, 2, 3, ..., I = 6 units) and the hidden-layer (j = 1, 2, 3, ..., J = 6 units), is represented by the summed value (Σz_i) at the output unit in Figure 1 . A hyperbolic tangent (sigmoidal) function, f(.), then squashes its linearly scaled value, K × Σz_i (where K is a linear-scaling constant), producing the result O_i = f(K × Σz_i). Additionally, the set {W_ij} specifies the coefficients of weighting of the connections between the input- and hidden-layers. Additionally, as shown, the ANN’s topology incorporates backpropagation and supervised learning enabled by a teacher value T_i. The gradient of an error, ε, which depends on (O_i, T_i), is defined by: (±ΔW_ij): and the entity α × (±ΔW_ij) is then applied iteratively to change the current value of W_ij until the error (ε) hits zero or a designated low, “stop” value. In this case, α stands for a learning coefficient that can be selected to attain a desired (quick) rate of convergence of the enforced repetition.

One effective method for training feed-forward neural networks is the backpropagation algorithm. However, it has a slow convergence rate and may produce less-than-ideal results because it updates the weights using the steepest descent approach [2] . As a result, a method that speeds up convergence is employed in this work [3] [4] . In other words, the authors have created a generalized process that, when the learning coefficient is chosen well, causes the ANN to converge to more accurate values more quickly [3] [4] .

The number of iterations needed to train the net is significantly reduced when this quicker method is used. The authors have demonstrated that the greatest eigenvalue of the Hessian matrix should be inversely proportional to the learning rate α [3] [4] . There are I = 6 input units, y = {y₁, y₂, …y_I} in Figure 1 . In the current example, y^Ty is an I × I (6 × 6) square matrix, and it corresponds to the Hessian matrix. This Hessian matrix can be expressed diagonally [HD]. The Hessian matrix’s symmetry results in a single, distinct eigenvalue, λ_II, in the diagonal form, as illustrated below (all other eigenvalues are zero):

$\left[HD\right]=\left[\begin{array}{ccccc}0& 0& 0& \dots & 0\\ 0& 0& 0& \dots & 0\\ 0& 0& 0& \dots & 0\\ \dots & \dots & \dots & \dots & \dots \\ 0& 0& 0& \dots & {\lambda }_{II}\end{array}\right]$ (1)

To facilitate faster convergence, the learning rate (α) applied to the test ANN would be equivalent to the reciprocal of the biggest, single eigenvalue, λ_II, of the Hessian matrix, as mentioned before. The input data for the current study relates to each ensemble of the test power load profile under investigation. The information provided by the Mazoon Electrical Company in Oman is the basis for the power load data that is taken into consideration. The meteorological department of Oman provided the temperature and humidity data. As a result, the effectiveness of the research conducted here using the suggested ANN-based approach and the power load predictions are cross-checked against specifics of the actual power load values.

This research predicts a monthly necessary power demand using the suggested neural network model and compares it with actual data. Using the learning rate as the reciprocal of the greatest eigenvalue of the Hessian matrix, it will be demonstrated that convergence towards accurate prediction is achieved rapidly.

An encouraging use of ANN for load forecasting has been shown in the current study. The outcomes, which are displayed in Table 4 , are extremely encouraging. The data confirms that the ANN predicts future power loads with a high degree of accuracy. Table 4 shows that the ANN prediction of power load closely matches the historical measured values.

A list of the study’s main findings is as follows: