Forecasting serves as a critical cornerstone for strategic planning, operational efficiency, and risk mitigation across modern civilization. By converting historical data into actionable forward-looking insights, it enables organizations and governments to anticipate market shifts, optimize resource distribution, and safeguard against systemic uncertainties. Predictive modeling is a fundamental task in many fields, such as finance, economics, engineering, and artificial intelligence. The aim of this paper is to summarize the methods of statistics and machine learning, outline their inherent challenges, and project future research directions. This paper mainly discusses traditional statistical methods (including Autoregressive Integrated Moving Average [ARIMA] and regression analysis), machine learning approaches (such as Random Forest and Support Vector Machines [SVM]), and deep learning models (such as Long Short-Term Memory [LSTM] networks and hybrid time series-ML models). Nowadays, as these interconnected fields become increasingly complicated, practitioners face severe challenges regarding data quality, computational complexity, and mathematical interpretability. This paper comprehensively reviews these methodologies, establishes a comparative taxonomy, and delineates the evolutionary trajectory of future forecasting applications.

Aiming at the problem that the original AlexNet convolutional neural network suffers from massive parameters and high floating-point computational overhead, which hinders real-time deployment on the Zynq FPGA embedded platform, this paper proposes a joint model compression scheme combining structured channel pruning and KL-divergence-based INT8 quantization. First, L1 regularization is applied for sparse training; 50 % low-contribution convolutional channels are pruned, followed by fine-tuning to restore model accuracy. KL-divergence calibration is then adopted to implement 8-bit integer quantization, which lightweight the model in terms of both computation and storage. Model training and compression are completed with PyTorch, while inference simulation and FPGA deployment are realized via TensorRT and Vivado. Ablation and comparative experiments are carried out on three datasets: COVID-19 chest X-ray images, the German Traffic Sign Recognition Benchmark (GTSRB), and CIFAR-10. Furthermore, dual visualization including Grad-CAM heatmaps and overlaid inference plots is adopted to analyze how lightweight operations affect feature extraction. Experimental results show that the joint pruning-plus-quantization scheme reduces model parameters from 60 M to 28 M, compresses model volume by 88 %, shortens single-image inference latency from 2.40 ms to 0.72 ms with a 3.3-fold speedup, and only brings 0.54 % classification accuracy drop. Visualization demonstrates that the lightweight model can accurately locate target objects with only minor degradation in fine-texture recognition. Under controllable accuracy loss, the proposed scheme significantly cuts hardware resource consumption and provides a practical lightweight solution for portable FPGA-based image-recognition systems in medical and vehicular scenarios.

In the current research, we developed a new method of pharmacodynamic data fusion for predicting biased signaling pathways related to G protein-coupled receptors (GPCRs). The β2-adrenergic receptor in humans (ADRB2) was chosen as the targeted GPCR, while the stable cell line of HEK293T cells with ADRB2 overexpression was used as an assay system. Four pharmacodynamics parameters were considered for data collection, namely cAMP generation, β-arrestin2 recruitment, ERK1/2 activation, and cell viability. The set of 40 molecules was screened as ADRB2 ligands at eight concentrations in triplicate. The process of data processing included four steps sequentially: curve fitting via GraphPad Prism 10, preprocessing in Python 3.11, neural network training in PyTorch 2.2, and SHAP. According to the model outputs, this novel multimodal strategy could classify the ligands as belonging to one of the three classes of biased signaling: Gs-biased, β-arrestin-biased, and balanced. In comparison to single-modality cAMP-based and β-arrestin2-based models, the multimodal model outperformed by its ability to distinguish ligands by their biased signaling profiles, reaching accuracy of 88.7%, macro-F1 of 0.872, and ROC-AUC of 0.931. Selection of key features showed four important features: cAMP pEC50, β-arrestin2 Emax, ERK1/2 max, and pathway ratios. According to SHAP results, cAMP pEC50 and β-arrestin2 Emax were two most important features, having contribution of 21.8% and 19.6%, correspondingly.

High penetration of wind and photovoltaic generation is changing the frequency dynamic characteristics of modern power systems. When synchronous generators are replaced by inverter based resources, the equivalent system inertia decreases. After disturbances, frequency changes more rapidly, RoCoF increases, frequency nadir becomes lower, and recovery time becomes longer. This study focuses on virtual inertia deficiency and control optimization in high renewable penetration power systems. The IEEE 39 bus standard system is used to construct 0 percent, 30 percent, 50 percent, and 70 percent renewable penetration scenarios. MATPOWER, Python, and MATLAB/Simulink are combined for power flow calculation, controller parameter optimization, and time domain simulation. Three control strategies are compared, including no virtual inertia control, fixed parameter VSG control, and MPC optimized VSG control. The results show that when renewable penetration increases from 0 percent to 70 percent, equivalent inertia decreases from 4.85 s to 1.76 s, the maximum RoCoF increases from 0.31 Hz/s to 0.91 Hz/s, and the frequency nadir drops from 49.72 Hz to 49.43 Hz. MPC optimized control improves the frequency nadir to 49.61 Hz in the 70 percent renewable scenario and reduces storage power stress. These results indicate that virtual inertia allocation should be jointly optimized with renewable penetration, critical buses, and storage constraints. The framework provides a reproducible technical path for frequency security control in high renewable power systems.

As China's national carbon market expands, fluctuations in carbon prices increasingly affect corporate operating costs, cash flow, and financing conditions, creating more complex transition risks for green credit. Using CEA trading data from July 2021 to December 2025 and quarterly observations of listed firms in regulated industries, ARMA-GARCH is applied to identify abnormal carbon price shocks. The KMV model measures distance to default and expected default frequency. Two-way fixed effects, mediation analysis, and XGBoost-SHAP are then used to examine risk transmission. The results show rapid carbon price increases during 2023 and 2024, followed by high trading activity and a price decline in 2025, indicating clear shifts in market risk states. Sectoral expansion also increased differences in carbon exposure, allowance demand, and cost absorption capacity across firms. Carbon price shocks therefore need to be evaluated together with firm-level carbon exposure and financial vulnerability. Their effects can be transmitted to credit risk through carbon costs, operating cash flow, and financing constraints. These findings provide quantitative support for green credit pricing, carbon stress testing, and early warning of credit risk among carbon-intensive firms.

Stock price forecasting is fundamentally difficult because of market non-stationarity, high volatility, and the weak predictability of intraday returns. While kernel methods with random features provide a theoretically grounded framework for capturing non-linear relationships, they are typically applied with a single, globally fixed feature representation that ignores the fact that the generative process of prices varies across market regimes. This paper proposes the Regime-Adaptive Multi-Scale Random Fourier Features (RA-MSRFF) model, which combines two ideas: (i) a multi-scale Random Fourier Feature (RFF) mapping that approximates multiple radial-basis kernels at different bandwidths and feeds a ridge regressor, and (ii) a soft regime-gating mechanism in which a volatility-driven fuzzy membership decomposes each sample into overlapping regimes, each equipped with its own local ridge expert, and the final prediction is a convex fusion of the global multi-scale estimator and the regime-aware local estimators. We establish that the regime-mixed kernel remains positive definite, that the fused objective stays convex, and that soft gating controls regime-misassignment risk. On two representative A-share bank stocks (601166, 601398) using a strict chronological train/validation/test protocol, RA-MSRFF strictly dominates five baselines—Open-price anchor, Static MS-RFF, KNN, AdaBoost, and Previous-close—on RMSE, MAE, MAPE, and R².

Reliable channel state information supports coherent multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) receivers. Least-squares channel estimation is simple but retains pilot noise, whereas fixed-window discrete Fourier transform least-squares (DFT-LS) denoising depends on a prescribed delay window. This paper proposes an adaptive energy-threshold DFT-LS refinement for sparse MIMO-OFDM channels. Cumulative delay-domain energy estimates a median-based tail noise floor, while amplitude thresholding selects taps over the full search interval. This separation prevents delayed weak taps from removal when an early strong tap satisfies the cumulative-energy condition. In sparse block-fading simulations with orthogonal pilots and uncoded quadrature phase-shift keying (QPSK), the proposed method reduces normalized mean-square error (NMSE) by approximately 42.3%–59.7% relative to fixed-window DFT-LS over 0–30 dB and reduces bit-error rate by 21.1% at 20 dB and 18.6% at 30 dB. The estimator remains non-iterative and requires less prior information than known-sparsity orthogonal matching pursuit (OMP-K) and oracle linear minimum mean-square error (Oracle-LMMSE) references.

Generative AI is reshaping content production in the cultural and creative industries by shifting cultural material organization, script writing, exhibition interpretation, and cultural tourism communication from linear manual workflows to human-AI collaborative workflows. Focusing on productivity improvement and the boundary of creative labor substitution, this research constructs a three-group comparison framework involving human-only creation, AI-assisted creation, and fully AI-generated creation. Based on public cultural heritage data, cultural policy texts, and cultural tourism communication samples from 2018 to 2024, LLMs, RAG, Sentence-BERT semantic retrieval, expert scoring, and mixed-effects ANOVA are combined to measure generation quality, policy suitability, and substitution intensity across different cultural tasks. The results show that the average completion time of the AI-assisted group decreases from 44.2 min to 25.7 min, while the overall quality score increases from 82.6 to 84.1, indicating that human-AI collaboration can improve production efficiency while maintaining content quality. Although the fully AI-generated group achieves the lowest time cost, its policy suitability and cultural accuracy decline noticeably. Task complexity further affects substitution intensity. Cultural tourism copy and short-video scripts are more substitutable, while exhibition interpretation and policy explanation still rely on human judgment. These findings provide quantitative evidence for AI adoption, creative labor restructuring, and cultural policy governance.

This study develops and evaluates a four-stage (V1–V4) progressive curriculum reinforcement learning framework for strategic badminton rallies between two simulated mobile manipulators. Each agent combines a Clearpath Ridgeback omnidirectional base, a Franka Panda arm, and a racket end effector in NVIDIA Isaac Lab. The task is formulated as a finite-horizon Markov decision process and optimized with Proximal Policy Optimization. The curriculum progresses from fixed-base single-shot interception (V1), through dual-robot mobile rallies with staged serve noise (V2), asymmetric tactical placement (V3), and symmetric full-court placement with predicted-intercept rewards and a product-form difficulty schedule (V4). The reported learning curves indicate that each stage acquired the behavior emphasized by its reward and curriculum design: stable interception in V1, improved rally continuity in V2, lateral placement and pursuit in V3, and more difficult tactical targets in V4. These results support progressive curriculum design as a practical way to mitigate sparse rewards in this simulated task. However, the evidence is limited to one simulation configuration and descriptive training curves; physical-robot validation, repeated-seed analysis, ablation studies, and non-curricular baselines remain necessary. The framework should therefore be interpreted as a simulation proof of concept and a basis for future sim-to-real evaluation, richer shot-type control, and opponent-diverse self-play.

In this paper, the proposed collaboration-based teaching model with incorporation of AI-assisted media arts practice is considered within the scope of secondary STEAM education. Two intact classes of Grade 8 students enrolled at a secondary school conducted the experiment for eight weeks. The experimental class took part in the coordinated instruction in science, technology, and art subjects and performed an assignment called "Urban Environmental Data and Interactive Media Art Project," where they used BBC micro:bit sensors, Adobe Express with Firefly, p5.js Web Editor, Google Classroom, and Padlet. The control class did a traditional STEAM project related to the same subject and period of time. Creativity test, project assessment rubrics, collaboration and engagement questionnaires, observations in classrooms, learning diaries, and interviews were utilized for evidence collection. The experimental class showed greater increase in creative problem solving compared to the control class (14.83 ± 5.21 vs. 5.87 ± 4.69) with statistically significant group effect. Furthermore, their projects had higher marks for data-art integration, interactivity, and justifications of design decisions. The score for collaboration and engagement achieved 4.31 ± 0.36 and 4.24 ± 0.41, respectively.