With the rapid development of modern industry, people’s living pressures are gradually increasing, and an increasing number of individuals are affected by sleep disorders such as insomnia, hypersomnia, and sleep apnea syndrome. Many cardiovascular and psychiatric diseases are also closely related to sleep. Therefore, the early detection, accurate diagnosis, and treatment of sleep disorders an urgent research priority. Traditional manual sleep staging methods have many problems, such as being time-consuming and cumbersome, relying on expert experience, or being subjective. To address these issues, researchers have proposed multiple algorithmic strategies for sleep staging automation based on deep learning in recent years. This paper studies MASleepNet, a sleep staging neural network model that integrates multimodal deep features. This model takes multi-channel Polysomnography (PSG) signals (including EEG (Fpz-Cz, Pz-Oz), EOG, and EMG) as input and employs a multi-scale convolutional module to extract features at different time scales in parallel. It then adaptively weights and fuses the features from each modality using a channel-wise attention mechanism. The integrated temporal features are integrated into a Bidirectional Long Short-Term Memory (BiLSTM) sequence encoder, where an attention mechanism is introduced to identify key temporal segments. The final classification result is produced by the fully connected layer. The proposed model was experimentally evaluated on the Sleep-EDF dataset (consisting of two subsets, Sleep-EDF-78 and Sleep-EDF-20), achieving classification accuracies of 82.56% and 84.53% on the two subsets, respectively. These results demonstrate that deep models that integrate multimodal signals and an attention mechanism offer the possibility to enhance the efficiency of automatic sleep staging compared to cutting-edge methods.

(1) Background: Ultra-short-term photovoltaic (PV) power prediction is crucial for optimizing grid scheduling and enhancing energy utilization efficiency. Existing prediction methods face challenges of missing data, noise interference, and insufficient accuracy. (2) Methods: This study proposes a single-step hybrid neural network model integrating Temporal Convolutional Network (TCN), Temporal Shift Transformer (TST), and Bidirectional Gated Recurrent Unit (BiGRU) to achieve high-precision 15-minute-ahead PV power prediction, with a design aligned with symmetry principles. Data preprocessing uses Variational Mode Decomposition (VMD) and random forest interpolation to suppress noise and repair missing values. A symmetric parallel dual-branch feature extraction module is built: TCN-TST extracts local dynamics and long-term dependencies, while BiGRU captures global features. This symmetric structure matches the intra-day periodic symmetry of PV power (e.g., symmetric irradiance patterns around noon) and avoids bias from single-branch models. Tensor concatenation and an adaptive attention mechanism realize feature fusion and dynamic weighted output. (3) Results: Experiments on real data from a Xinjiang PV power station, with hyperparameter optimization (BiGRU units, activation function, TCN kernels, TST parameters), show that the model outperforms comparative models in MAE and R²—e.g., the MAE is 26.53% and 18.41% lower than that of TCN and Transforme. (4) Conclusions: The proposed method achieves a balance between accuracy and computational efficiency. It provides references for PV station operation, system scheduling, and grid stability.