**5. Prediction model—Artificial neural network**

The artificial neural network (ANN) is a popular AI model and a robust computational tool based on the human brain's organizational structure [13]. ANNs are the

**Figure 3.** *A schematic of the developed model [8].*

#### *Improve Energy Efficiency in Surface Mines Using Artificial Intelligence DOI: http://dx.doi.org/10.5772/intechopen.101493*

representation of methods that the brain uses for learning which are known as neural networks (NNs), simulated neural networks (SNNs), or parallel distributed processing (PDP). ANN simulates the effect of multiple variables on one significant parameter by a fitness function. Thus, ANNs are excellent solutions for complex problems as they can signify the compound relationships between the various parameters involved in a problem.

ANN methods are established as powerful techniques to solve various real-world problems among the different machine intelligence procedures due to ANN's excellent learning capacity in recent decades. The approximate solution by ANN is found to be useful, but it depends upon the ANN model that one considers [14].

Layers are commonly used to organize neural networks. Layers are made from various interconnected "neurons/nodes," which include "activation functions." ANN processes information to solve problems through neurons/nodes in a parallel manner. First, ANN obtains knowledge through learning and is stored within interneuron connections'strength, expressed by numerical values called "weights." Then, these weights and biases are combined to calculate output signal values for a new testing input signal value. Next, patterns are provided to the network through the "input layer," which connects to one or more "hidden layers," where the actual processing is completed through a system of weighted "connections." The hidden layers then correlate to an "output layer," which generates the output through the activation functions [Eqs. (1)–(3)].

$$E\_k = \sum\_{j=1}^{q} (w\_{ijk}\mathbf{x}\_j + b\_{ik}) \qquad \qquad k = \mathbf{1}, \ \mathbf{2}, \ \dots, \ m \tag{1}$$

Where *i* is the input, *x* is the normalized input variable, *w* is the weight of that variable, *b* is the bias, *q* is the number of input variables, and *k* is the counter of neural network nodes, and *m* is the number of neural network nodes in the hidden layer.

In general, the activation functions contain linear and nonlinear equations. The coefficients related to the hidden layer are grouped into matrices *wijk* and *bik*. Eq. (2) is often used as the activation function between the hidden and output layers, where *f* is the transfer function.

$$F\_k = f(E\_k) \tag{2}$$

The output layer calculates the weighted sum of the signals provided by the hidden layer, and the related coefficients are grouped into matrices *Wok* and *bo*. Thus, the network output can be determined by Eq. (3).

$$Output = \left(\sum\_{k=1}^{m} w\_{ok} F\_k \right) + b\_o \tag{3}$$

The most significant component of neural network modeling is network training, which can be done in two ways: controlled and uncontrolled. Backpropagation is the most widely used training algorithm, which was established after examining several types of algorithms. A training algorithm modifies the coefficients (weight and bias) of a network to reduce the error between the estimated and actual network outputs.

The Mean Square Error (MSE) and Coefficient of Determination (R<sup>2</sup> ) were used in this study to investigate the error and performance of the neural network output and determine the appropriate number of nodes in the hidden layer. **Figure 4** depicts the created model's basic structure.

#### **Figure 4.**

*Structure of artificial neural network [8].*


#### **Table 1.**

*Case studies information.*

The developed AI model was tested against actual data taken from standard trucks in two surface mines in Australia and Iran. **Table 1** contains some information from these case studies.

For a standard range of loads, **Figures 5** and **6** show the correlation between P, S, T.R., and FCIndex created by the constructed ANN model for two types of standard trucks employed in case studies.

The presented graphs show a nonlinear relationship between FCIndex and P. The fuel consumption rate increases dramatically with increasing T.R. However, this rate does not change sharply with changing truck speed (S).

The results show good agreement between the estimated and actual values of fuel consumption. **Figure 7** presents sample values for the independent (tested) and the estimated (using the ANN) fuel consumption to highlight the insignificance of the importance of the absolute errors in the analysis for studied mines.

*Improve Energy Efficiency in Surface Mines Using Artificial Intelligence DOI: http://dx.doi.org/10.5772/intechopen.101493*

**Figure 5.**

*Correlation between payload, S,T.R., and FCIndex based on the developed ANN model for CAT 793D (mine 1).*

**Figure 6.**

*Correlation between GVW, S,T.R., and FCIndex based on the developed ANN model for Komatsu HD785 (mine 2).*

## **6. Optimization model—Genetic algorithm**

Optimization is a branch of computational science that shows how to find the best measurable solution to various issues. It is critical to consider the search area and goal function components when solving a specific problem. All the solution's possibilities are investigated in the search area. The objective function is a mathematical function that connects each point in the search area to an actual value that may be used to evaluate all search area members.

Traditional optimization methods are described by the stiffness of their mathematical models and limit their application in presenting dynamic and complex

#### **Figure 7.**

*Sample values for the estimated and the independent fuel consumption index.*

situations of "real life." Optimization techniques based on AI, underpinned by heuristic rulings, can reduce the problem of stiffness and are suitable to solve various kinds of engineering problems.

Some heuristic algorithms were developed in the 1950s to replicate biological processes in engineering. When computers were developed in the 1980s, it became possible to employ these algorithms to optimize functions and processes, whereas older methods failed.

During the 1990s, some new heuristic methods were developed by prior algorithms, such as Swarm Algorithms, Simulated Annealing, Ant Colony Optimization, and (GA). GA is one of the most widely used evolutionary optimization algorithms.

GAs were proposed by Holland (1975) based on an abstraction of biological evolution using ideas from natural evolution and genetics to design and implement robust adaptive systems [15]. In optimization methods using the new generation of GA is relatively novel. Moreover, they have good chances to escape from local minimums because of no need for any derivative information. As a result, their application in practical engineering problems can provide more satisfactory solutions than other traditional mathematical methods [16].

GAs are similar to the evolutionary aspects of natural genetics. The individuals are randomly selected from the search area. The fitness of the solutions is determined from the fitness function, subsequently. It is the result of the variable that is to be optimized. The individual that creates the best fitness in the population (a group of possible solutions) has the highest chance to return in the next generation with the opportunity of reproduction by the crossover with another individual, thus producing decedents with both characteristics. The possible solutions will converge to an optimal solution for the proposed problem by correctly developing a GA Crossover, which contributes to the evolution based on selection, reproduction, and mutation.

Due to their potential as optimization techniques for complex functions, GAs have been used in various scientific, engineering, and economic problems [17–20]. There are four significant advantages of using GAs to optimize problems [21]:

*Improve Energy Efficiency in Surface Mines Using Artificial Intelligence DOI: http://dx.doi.org/10.5772/intechopen.101493*


It is crucial to investigate the impact of particular parameters on GA behavior and performance to determine their relevance to the problem requirements and available resources. Furthermore, the type of problem being addressed determines the impact of each parameter on the algorithm's performance. As a result, determining the best values for these characteristics will necessitate a significant amount of experimentation.

In the GA model, Fitness Function, Individuals, Populations and Generations, Fitness Value, Parents and Children are the main parameters [17]. In addition, the population size impacts global performance and GA efficiency, and the mutation rate ensure that a given position does not remain fixed in value or the search becomes essentially random. **Figure 8** depicts the basic framework of a GA model.

A GA model was created to optimize the significant, influential factors on the energy consumption of haul trucks. **Tables 2** and **3** show the outcomes of utilizing the proposed model for actual case studies with an optimal range of variables.

Using the developed AI models in the two studied mines site shows energy efficiency improvements between 9 and 12%. Reaching the mentioned fuel consumption reduction and energy efficiency is promising when one mostly used truck in the mine site consumes around 110 L of diesel per hour. The haul trucks normally are used 24 h and 7 days per week to move mined materials in the site. Studied mine site had more than 100 trucks in their fleet, and the average price of diesel in those regions was 1.3 dollars per liter. It means that 9–12% energy efficiency improvement equals millions of dollars in saving annually.

**Figure 8.** *GA processes (developed model) [8].*


**Table 2.**

*The result of the GA model for CAT 793D in mine (1).*


**Table 3.**

*The result of the GA model for HD 785 in mine (2).*
