Journal of Artificial Intelligence and Metaheuristics

With the growing complexity of energy consumption patterns, accurate forecasting is required to maximize the benefits of demand-side management, which relies on precise forecasting and optimization. To respond to this challenge, this research suggests a new forecasting method by integrating the Al-Biruni Earth Radius (BER) optimization algorithm with the Long Short-Term Memory (LSTM) network within an Internet of Things (IoT)-based framework. The BER-optimized LSTM model processes real-time IoT sensor data, such as energy usage patterns and weather conditions, to enhance energy usage predictions. Time-series clustering addresses data uncertainty and noise, providing a probabilistic load forecast at both the household and aggregated levels. Extensive evaluation of the model’s performance demonstrates substantial predictive accuracy, achieving a root mean squared error (RMSE) of 0.007146, surpassing conventional forecasting methods. These findings corroborate that BER optimization, coupled with the use of LSTM networks, provides a suitable tool for smart home energy forecasting, which, in turn, can facilitate improved energy management, demand-side optimization, and innovative grid applications in the evolving energy market.

Time series forecasting is crucial in agricultural markets, where accurate predictions can enhance stakeholder decision-making. Traditional statistical models such as ARIMA often struggle with capturing complex patterns, necessitating the integration of metaheuristic optimization techniques to improve predictive performance. This study introduces a novel hybrid optimizer, the Grey Wolf-Greylag Goose Optimizer (GGGWO), which combines the strengths of the Grey Wolf Optimizer (GWO) and the Greylag Goose Optimizer (GGO) to enhance the forecasting accuracy of ARIMA models. The proposed GGGWO-ARIMA model is evaluated on a dataset of daily potato prices across multiple Indian cities and is compared against ARIMA, WOA-ARIMA, PSO-ARIMA, GWO-ARIMA, and GGO-ARIMA. Experimental results demonstrate that GGGWO achieves the lowest MSE (0.0027), RMSE (0.0184), and MAE (0.0116) while attaining the highest R-squared (0.9648) and Willmott Index (0.9565), outperforming all baseline models. These findings highlight the efficacy of hybridizing GWO and GGO, offering a robust optimization framework for improving time series forecasting in agricultural price prediction. This can aid policymakers, farmers, and market analysts make data-driven decisions.

Pneumonia diagnosis from chest X-ray images remains clinically important because subtle radiographic abnormalities can be difficult to assess consistently under high workload or limited expert availability. This study proposes an optimization-guided diagnostic framework that integrates the Ocotillo Optimization Algorithm (OcOA) with a Convolutional Neural Network (CNN) for binary pneumonia classification. Rather than introducing OcOA or CNN as new algorithms, the contribution lies in using OcOA in two complementary stages: feature selection over CNN-derived deep representations and hyperparameter optimization of the CNN classifier. The proposed pipeline was evaluated on the Chest X-Ray Images (Pneumonia) dataset, which contains pediatric radiographs labeled as Normal or Pneumonia. In the feature-selection stage, binary Ocotillo Optimization Algorithm (bOcOA) achieved the lowest average error, with a value of 0.523417. After feature selection, the CNN baseline achieved the best classification performance among the evaluated machine-learning models, reaching an accuracy of 0.94985. In the hyperparameter optimization stage, OcOA+CNN achieved the strongest optimized performance, with an accuracy of 0.989939. These findings indicate that OcOA can improve CNN-based pneumonia detection by refining both the feature space and the model configuration. The framework may support automated chest X-ray screening on the studied dataset; however, external validation, statistical testing, and prospective clinical assessment remain necessary before any clinical deployment claim can be made.

Accurate drug classification is important for developing intelligent healthcare systems and supporting reliable medication-related decision-making. The relationships between patient characteristics and prescribed drug categories may be nonlinear and difficult to identify using conventional analytical procedures. This study presents a machine learning framework for multiclass drug classification using demographic and clinical patient attributes. The framework includes data cleaning, exploratory data analysis, numerical and categorical feature transformation, model training, comparative evaluation, and explainable artificial intelligence analysis. Nine classification models were investigated, including Random Forest, support vector machine, LightGBM, CatBoost, multilayer perceptron, one-dimensional convolutional neural network, long short-term memory network, gated recurrent unit network, and TabNet. The models were evaluated using accuracy, weighted precision, weighted recall, weighted F1-score, macro F1-score, and balanced accuracy. Random Forest achieved the best overall balance between predictive performance and interpretability, obtaining an accuracy of 97.50%, a macro F1-score of 98.51%, and a balanced accuracy of 98.18%. SHAP analysis identified the sodium-to-potassium ratio, blood-pressure categories, and age as the most influential variables affecting the classification decisions. The obtained results demonstrate that an explainable machine learning framework can provide accurate and transparent drug classification using a limited number of patient-related variables.

Optimization plays a fundamental role in engineering design, enabling cost reduction, performance enhancement, and constraint satisfaction. Metaheuristic algorithms such as the Grey Wolf Optimizer (GWO) and Dipper-Throated Optimizer (DTO) have been widely used for solving complex optimization problems. However, standalone algorithms often suffer from premature convergence and limited exploration capabilities, necessitating the development of hybrid approaches. This chapter introduces a novel hybrid algorithm, GWO+DTO, which combines the exploratory strength of GWO with the exploitative efficiency of DTO to improve optimization performance. The effectiveness of GWO+DTO is evaluated on two benchmark engineering problems: the Pressure Vessel Design Problem and the Tension/Compression Spring Design Problem, comparing its results with standalone GWO and DTO. Experimental findings demonstrate that the hybrid approach achieves superior performance, obtaining the best cost of 5950.28 in the pressure vessel problem and 0.01266 in the spring design problem, outperforming the individual algorithms in accuracy and efficiency. Additionally, GWO+DTO requires fewer function evaluations, highlighting its computational efficiency. The proposed hybrid method presents a promising alternative for tackling real-world engineering optimization challenges, with potential applications in multi-objective and large-scale optimization problems.