Accurate base station traffic data in a public place with large changes in the amount of people could help predict the occurrence of network congestion, which would allow us to effectively allocate network resources. This is of great significance for festival network support, routine maintenance, and resource scheduling. However, there are a few related reports on base station traffic prediction, especially base station traffic prediction in public scenes with fluctuations in people flow. This study proposes a public scene traffic data prediction method, which is based on a $v$ Support Vector Regression (vSVR) algorithm. To achieve optimal prediction of traffic, a symbiotic organisms search (SOS) was adopted to optimize the vSVR parameters. Meanwhile, the optimal input time step was determined through a large number of experiments. Experimental data was obtained at the base station of Huainan Wanda Plaza, in the Anhui province of China, for three months, with the granularity being one hour. To verify the predictive performance of vSVR, the classic regression algorithm extreme learning machine (ELM) and variational Bayesian Linear Regression (vBLR) were used. Their optimal prediction results were compared with vSVR predictions. Experimental results show that the prediction results from SOS-vSVR were the best. Outcomes of this study could provide guidance for preventing network congestion and improving the user experience.

Rotor attitude estimation (RAE) is a predominant approach to controlling spherical motors, but there is a margin for its improvement. This article proposes a visual RAE method by using the Kalman filter-based multi-object fast discriminative scale space tracker (KMFDSST) algorithm. The KMFDSST algorithm is adopted to detect three visual objects simultaneously on the top of the spherical motor. The rotor attitude is estimated based on the positions of the three objects. To verify the accuracy and dynamic performance of the KMFDSST algorithm when the occlusion cases happened at large tilt angles, the one-object tracking simulations are conducted among the KMFDSST, fast discriminative scale space tracker (FDSST), and multi-object Kalman kernelized correlation filter (MKKCF) algorithms. Simulation and experiment results indicate that the robustness of the KMFDSST algorithm is better than that of both MKKCF and FDSST algorithms. Moreover, the comparative experiment between the KMFDSST and micro-electro mechanical system (MEMS) RAE methods shows the advantages of the proposed RAE method.

Accurate and efficient prediction of mobile network traffic in a public setting with changing flow of people can not only ensure a stable network but also help operators make resource scheduling decisions before reasonably allocating resources. Therefore, this paper proposes a method based on kernel extreme learning machine (kELM) for traffic data prediction. Particle swarm optimization (PSO), multiverse optimizer (MVO), and moth–flame optimization (MFO) were adopted to optimize kELM parameters for finding the best solution. To verify the predictive performance of the kernel ELM model, backpropagation (BP) neural network, v-support vector regression (vSVR), and ELM were also applied to traffic prediction, and the results were compared with kELM. Experimental results showed that the smallest mean absolute percentage error in the test (11.150%) was achieved when kELM was optimized by MFO with Gaussian as the kernel function, that is, the prediction result of MFO-kELM was the best. This study can provide significant guidance for network stability and resource conservation.

An HPO (Hunter–Prey Optimizer)-optimized Bidirectional LSTM (HPO-BiLSTM) model is introduced to address the challenges in predicting gas concentration within coal mining working faces. This study aims to adaptively adjust the key hyperparameters (such as learning rate and number of hidden layer units) of the BiLSTM network through intelligent optimization algorithms. While the BiLSTM architecture inherently mitigates gradient vanishing and exploding problems through its gating mechanisms, the proposed HPO method focuses on addressing the inefficiency of manual parameter tuning and the risk of trapping in local optima that traditional methods encounter when dealing with nonlinear and non-stationary gas concentration time series. The experiment utilized the actual methane monitoring data from the 15117 working face of Jishazhuang Coal Mine in Jinzhong City, Shanxi Province (with a sampling interval of 2 min). The proposed HPO-BiLSTM model was compared with baseline models such as LSTM, BiLSTM, GA-BiLSTM, and PSO-BiLSTM in terms of performance. This study systematically compares the performance of LSTM, BiLSTM, and BiLSTM models optimized with GA, PSO, and HPO. Results demonstrate that all optimized models outperform the baselines, with HPO-BiLSTM achieving the best overall performance. It attained the lowest RMSE and highest R² across the training, validation, and test sets, showcasing superior fitting and generalization capabilities. Furthermore, HPO-BiLSTM converged to the lowest loss value (0.00062) in only 15 iterations, demonstrating significantly greater efficiency and stability than both GA-BiLSTM (loss 0.00072, 25 iterations) and PSO-BiLSTM (loss 0.00071, 30 iterations). The experiments confirm that the HPO algorithm effectively configures BiLSTM hyperparameters, mitigates overfitting, and provides a more accurate and robust solution for gas concentration prediction in coal mines.