The ever-increasing widespread use of the Internet of Things and its applications has generated massive amounts of data. IoT sensor-generated datasets typically have a time-series structure and relational metadata to describe them. Time series data are typically data that have timestamps and can be obtained from sensors and IoT devices. Data prediction is required to maximize the potential of IoT-generated data, and anomaly detection and correction are needed to preserve data quality and integrity. Traditional machine learning models are incapable of analyzing the gigantic amounts of IoT-generated data. On the other hand, deep learning can properly analyze large data volumes, leading to increased use in the IoT domain. This research has examined the use of deep learning models for prediction, anomaly detection, and correction of data generated by IoT devices. The study found that deep learning is widely used in different fields today to analyze IoT-generated data. The research also outlines some challenges being faced while using deep learning models for IoT data analysis. More research is suggested in this study to expose more challenges and tackle the current challenges to achieve better IoT data analysis using deep learning.

This study re-examines the balance between architectural intricacy and generalization in Transformer models for long-term time series predictions. We perform a systematic comparison involving a lightweight baseline (iTransformer) and two enhanced versions: MiTransformer, which incorporates an external memory component for extending context, and DFiTransformer, which features dual-frequency decomposition along with Learnable Cross-Frequency Attention. All models undergo training using the same protocols across eight standard benchmarks and four forecasting periods. Findings indicate that both MiTransformer and DFiTransformer do not reliably surpass the baseline. In many instances, the increased complexity leads to greater variance and decreased accuracy, especially with unstable or inconsistent datasets. These results imply that architectural minimalism, when effectively refined, can match or surpass the effectiveness of more complex designs—challenging the prevailing trend toward increasingly intricate forecasting architectures.

Transformer-based models for time series forecasting have advanced considerably, yet many approaches treat attention sparsity and memory mechanisms as separate strategies. In this extended work, we revisit these paradigms and present LTSMiTransformer, a unified architecture that integrates Learnable Temporal Sparse Attention (LTSA), a Memory-Augmented Module (MAM) and a unified embedding strategy that enhances feature representation across heterogeneous datasets to improve efficiency, scalability, and generalization in long-horizon multivariate time series forecasting. LTSA employs a trainable threshold to dynamically filter irrelevant time steps, reducing attention complexity from O (L²) to O (L log L), while MAM encodes long-term patterns into a compact memory bank with gated updates, enabling persistent context retention without uncontrolled memory growth. Unlike prior models, these components are jointly optimized to balance sparsity and memory in a coherent training framework. This extended work provides new theoretical analysis of LTSA’s sub-quadratic efficiency and MAM’s convergence behaviour under sparse gating. Comprehensive robustness tests under noise, missing values, and temporal imbalance, along with evaluations on eight diverse datasets from finance, energy, weather, and traffic, demonstrate LTSMiTransformer’s accuracy, generalizability, and compactness. The theoretical proofs, ablation studies, and robustness evaluations included show that the LTSMiTransformer achieves a strong balance between scalability, interpretability, and generalization across diverse datasets. This paper significantly expands our earlier CNIOT 2025 conference papers by adding new theoretical results, deeper joint-training analysis, and wider empirical validation. All source code and datasets are available on GitHub to support reproducibility.

According to recent time series forecasting research, simpler models frequently perform better than their more complex counterparts, particularly over longer time horizons. By comparing two improved versions of the iTransformer—MiTransformer, which incorporates an external memory module, and DFiTransformer, which adds dual-frequency decomposition and Learnable Cross-Frequency Attention—this study investigates this assertion. Although both seek to increase forecasting accuracy, empirical findings across a number of benchmarks demonstrate that performance is frequently negatively impacted by complexity. The conclusion that simpler, well-structured architectures can provide superior generalisation and practical utility is further supported by the notable underperformance of the most complex model, DFiTransformer.