Tidal time series are affected by a combination of astronomical, geological, meteorological, and anthropogenic factors, revealing non-stationary and multi-period features. The statistical features of non-stationary data vary over time, making it challenging for typical time series forecasting models to capture their dynamism. To solve this challenge, we designed memory factors, leveraging the fusion of statistical data at the channel dimension to enhance the model's prediction capacity for non-stationary data. On the other hand, traditional approaches have limitations in trend and cycle decomposition, making it difficult to detect complicated multi-period patterns and accurately separate the components. We combined integrated frequency domain optimization and multi-level, multi-scale convolutional kernel technologies. By employing Fourier-based methods and iterative recursive decomposition strategies, we effectively separated periodic and trend components. Then, the periodic multi-level wavelet block was applied to extract the periodic interaction features, aiming to deeply mine the latent information of periodic components and enhance the model's long-term prediction capabilities. In this paper, we used the Informer model as the foundational framework for further research and development. In comparative experiments, our proposed model outperformed LSTM, Informer, and MICN by 61.4%, 51.7%, and 23.8%, respectively. In multi-time-span prediction, the model's error remained stable as the prediction span increased from 48 to 96 steps (from 0.059 to 0.067). Under multi-site conditions, the model achieved varying degrees of improvement over the baseline in three key evaluation metrics, with average increases of 35.2%, 35.6%, and 61.2%, respectively. In this study, we focused on the extraction of short-period features from tidal data, providing an innovative and reliable solution for tidal height prediction. The results are significant for tidal assessments and protective engineering construction.

Wave height prediction is hampered by the volatility and unpredictability of ocean data. Traditional single predictors are inadequate in capturing this complexity, and weighted fusion methods fail to consider inter-model correlations, resulting in suboptimal performance. To overcome these challenges, we presented an improved stacking-based model that combined the long short-term memory (LSTM) network with extremely randomized trees (ET) for wave height prediction. Initially, features with weak correlation to wave height were excluded using the Pearson correlation coefficient. Subsequently, a stacking ensemble tailored for time series cross-validation was deployed, employing LSTM and ET as base learners to capture temporal and feature-specific patterns, respectively. Lasso regression was utilized as the meta-learner, harmonizing these insights to improve accuracy by leveraging the strengths of each model across different dimensions of the data. Validation using datasets from four buoy stations demonstrated the superior predictive capability of our proposed model over single predictors such as temporal convolutional networks (TCN) and XGBoost, and fusion methods like LSTM-ET-BP.