Frontiers in Computing and Intelligent Systems ISSN: 2832-6024 | Vol. 12, No. 1, 2025 35 TimeMixer‐ME: A Dual‐Memory Enhanced Architecture for Time Series Forecasting Dongze Wu, Yonghai Yang, Dechuan Qiu *, Guangming Li University of Science and Technology West Campus, 89 Beijing Road, Korla, Bayingolin Mongol Autonomous Prefecture 841000, Xinjiang, China * Corresponding author: Dechuan Qiu (Email: dimitriculter@foxmail.com) Abstract: Time series prediction, as a research field with extensive practical applications, consistently faces a key challenge: how to improve the model's overall predictive capability while maintaining short-term prediction accuracy. To address this challenge, this research innovatively proposes TimeMixer-ME, a dual-memory enhanced architecture that ingeniously integrates cognitive science's dual memory system theory with deep learning techniques, achieving adaptive multi-scale modeling of time series. Through experimental validation on multiple weather datasets from Southern Xinjiang, TimeMixer-ME demonstrates exceptional predictive performance. Compared to the baseline TimeMixer model, it achieved a significant reduction of 2%-2.4% in MSE metrics and a remarkable improvement of 3.7%-4.2% in MAE metrics. These results convincingly demonstrate the model's excellent performance in short-term prediction tasks, providing a novel approach to addressing the trade-off between long-term and short-term dependencies in time series prediction. Keywords: Dual Memory-enhanced Architecture; Cognitive Science; Adaptive Multi-scale Modeling; Short-term Prediction. 1. Introduction In recent years, the southern region of Xinjiang has experienced a continuous increase in annual average precipitation and frequent extreme weather events, significantly impacting local agricultural production and residents' lives. Accurate short-term precipitation forecasting is crucial for disaster prevention. However, due to the unique geographical environment and climatic characteristics of Southern Xinjiang, existing precipitation forecasting models face multiple challenges in this region: Located between the Tianshan and Kunlun Mountains, the diverse topographical features including mountains, basins, and gobi desert lead to complex local weather systems. Existing models perform poorly in handling mountain-plain circulation and local convection systems. Furthermore, meteorological elements such as temperature, humidity, pressure, and wind fields exhibit unique spatiotemporal variation characteristics and interaction patterns during precipitation formation, increasing the difficulty of capturing coupling relationships. Additionally, due to terrain and observation station distribution limitations, meteorological observation data in this region suffers from uneven spatial distribution and discontinuous time series. The region's precipitation shows significant seasonal variation characteristics, with substantial differences in precipitation mechanisms across seasons, collectively making it difficult for existing forecasting models to maintain stable prediction accuracy. To address these challenges, this paper proposes a novel dual-memory enhanced architecture, TimeMixer-ME , based on TimeMixer. The model achieves innovative breakthroughs in multiple aspects: Designed for Southern Xinjiang's characteristics, it incorporates a dual-memory system with short-term and long-term memory units. The short-term memory unit employs 3×1 depth-separable convolution and 4-head attention mechanism to capture rapid changes in local weather systems, while the long-term memory unit utilizes a 32-dimensional memory bank and adaptive pooling mechanism to store and leverage seasonal climate characteristics, enabling effective modeling of multi-scale weather processes. The model innovatively introduces an adaptive feature fusion mechanism, dynamically adjusting the importance of different meteorological elements through learnable weight matrices, enhancing understanding and prediction capabilities of complex weather system evolution. Furthermore, through specialized data preprocessing and augmentation strategies, and the introduction of a periodic detector based on 7×1 convolution kernels, the model significantly improves its robustness to incomplete and noisy data. Experiments show that compared to the baseline TimeMixer model, TimeMixer-ME achieves 2%-2.4% lower MSE and 3.7%-4.2% lower MAE on multiple Southern Xinjiang weather datasets, demonstrating excellent short- term prediction accuracy. These results confirm the practical value of the proposed model in short-term precipitation forecasting for the Southern Xinjiang region. The remainder of this paper is organized as follows: Section 2 reviews related work; Section 3 details the TimeMixer-ME model architecture; Section 4 presents experimental results and analysis in Southern Xinjiang; and finally, Section 5 concludes the paper and discusses future improvement directions. 2. Related Work Research methods in time series forecasting have undergone significant evolution from traditional statistical approaches to deep learning methods. Early research was represented by the ARIMA model proposed by Box et al.[1] and systematic forecasting principles by Hyndman[2], which established a solid theoretical foundation for time series analysis. The advent of deep learning brought revolutionary breakthroughs to the field, with Graves et al.[3] proposing LSTM, which effectively addressed long-term dependency problems through sophisticated gating mechanisms, while the Transformer architecture proposed by Vaswani et al.[4] 36 initiated a new research wave with its self-attention mechanism.Building upon these foundations, FEDformer proposed by Zhou et al.[5] enhanced periodic pattern modeling capabilities through frequency domain decomposition, while Wu et al.[6] systematically validated the effectiveness of Transformer architectures in time series forecasting. Addressing non-stationarity issues, Liu et al.[7] introduced the Non-stationary Transformer with innovative adaptive normalization mechanisms.In terms of multi-scale modeling, TimesNet by Wu et al.[8] achieved effective multi- scale feature capture through temporal 2D variation modeling, while Scaleformer proposed by Shabani et al.[9] further improved prediction accuracy using iterative refinement strategies. Recent research has focused more on quantifying prediction uncertainty, with Lin et al.[10] providing a comprehensive review of diffusion models in time series applications, while the Koopa architecture proposed by Liu et al.[11] innovatively incorporated Koopman operator theory.Regarding memory enhancement mechanisms, Lee et al.[12] and Liu et al.[13] proposed pattern matching memory networks for traffic prediction and graph learning memory networks for multivariate time series prediction, respectively. Additionally, Crossformer proposed by Zhang et al.[14] provided new solutions for multivariate time series prediction by leveraging cross-dimensional dependencies. 3. TimeMixer-ME Input series DFT Series Decomposition Frequency Domain Decomposition MultiScaleSeasonMixing Bottom-up-mixing MultiScaleTrendMixing Up-bottom-mixing AnyVariateAttention Multi-Timescale Memory Enhancement Mechanism (MTMEM) Short_term _memory Central memory Long_term _memory Result Figure 1. The Overall Architecture of TimeMixer-ME We extend the baseline TimeMixer model which includes DFT Series Decomposition and MultiScaleMixing architecture modules by introducing the AnyVariateAttention module and proposing the MTMEM module. 3.1. AnyVariateAttention Input LayerNorm K‐>K(2D) V‐>V(D) Q‐>Q(2D) V K Q Scaled dot- product attention + u1/u2 Softmax Dropout Output_proj ection +Residual Output Figure 2. The architectural diagram of the AnyVariateAttention module The input features first undergo LayerNorm preprocessing, followed by parallel processing through three branches to generate Query (Q), Key (K), and Value (V) representations. These representations are then transformed into their respective 2D formats (Q->Q(2D), K->K(2D), V->V(D)). The Key and Query branches feed into a scaled dot-product attention mechanism, which is enhanced by a u1/u2 scaling factor before passing through Softmax and Dropout layers. The attention output is combined with the processed Value branch in the Output_projection layer, and finally merged with a residual connection to produce the final output. The complex interactions between multivariate time series exhibit unique coupling characteristics across different temporal and feature dimensions. Given an input sequence: ∈ 1 where B represents batch size, T denotes sequence length, N indicates the number of variables, and D is the feature dimension, AnyVariateAttention enhances inter-variable feature representation through the following integrated mechanism:The module first projects input features into a 37 higher-dimensional representation space to better capture multi-scale feature relationships, where: ∈ 2 ∈ 3 ∈ 4 This double-dimension projection design significantly enhances the model's expressiveness. To strengthen positional awareness, a learnable rotary position encoding matrix R is introduced, where: , ∈ 5 which adaptively learns positional dependencies through gradient descent. The attention computation considers inter- variable relationships through: √2 ∈ 6 incorporating variable-specific bias terms: 1 7 to distinguish relationships between same and different variables, resulting in: 8 where and are learnable scalar parameters and is the N-dimensional identity matrix. The final attention weights are computed through: softmax ∈ 9 followed by feature aggregation: ∈ 10 Layer normalization and residual connections ensure feature distribution stability through LayerNorm 11 and 12 where is the output projection matrix. The complete forward propagation process can be expressed as: AnyVariateAttention LayerNorm ⋅ 13 Where: softmax √2 1 14 This design achieves an efficient balance between computational complexity and model expressiveness. Through adaptive variable attention mechanisms, the model effectively captures dynamic correlations in multivariate time series while differentiating between same-variable and cross-variable relationships through bias terms and . The temporal structure is preserved through rotary position encoding and residual connections, while learnable projection matrices and layer normalization enable adaptive feature representation adjustment, significantly enhancing the model's capability in complex multivariate time series modeling while maintaining computational efficiency. 3.2. Multi-Time-scale Memory Enhancement Module Input w ₁ w ₂ w ₃ w ₄ w ₅ w ₆ w ₇ 7 1 Conv Extracted periodic features G E L U Enhanced periodicity features 3×1 local convolution w ₁ w ₂ w ₃ S to ra ge U pd at e R et ri ev al Short_term memory unit Q K V Four_Head attention Enhanced short-term dependency features GAP S to ra ge U pd at e R et ri ev a Long_term memory unit Enhanced long-term dependency features Periodicity detector Short_term m em ory unit L ong_term m em ory unit Feature integration LayerNorm Dropout Result Figure 3. The architecture of MTMEM This module consists of three parallel processing tracks for comprehensive feature extraction and enhancement. The upper track employs GAP (Global Average Pooling) followed by a long-term memory unit to capture global dependencies. The middle track utilizes a 3×1 local convolution and short- term memory unit, coupled with Four-Head attention mechanism for local feature processing. The bottom track implements a 7×1 convolutional layer and GELU activation to extract periodic features. Each track enhances specific feature types (long-term dependency, short-term dependency, and periodicity) before being integrated through a feature integration module. The final output undergoes LayerNorm and Dropout operations to produce the result. Time series contain rich multi-scale temporal features that exhibit unique statistical properties at different time scales. Given the input sequence: ∈ 15 the Multi-Time-scale Memory Enhancement Module (MTMEM) enhances time series representation through a multi-level feature extraction and fusion mechanism, expressed as: , ℓ , 16 38 The core architecture of MTMEM consists of three complementary feature extraction modules, each focusing on capturing temporal patterns at specific scales. First, short-term dependency features are captured through a combination of local convolution and attention mechanisms. Specifically, depthwise separable convolution extracts features within a local time window as: ∗ , ∈ 17 This convolution structure significantly reduces the number of parameters while maintaining feature extraction capability. Subsequently, a multi-head self-attention mechanism further enhances the expression of local features: softmax 18 where the attention weights consider the local correlations of features, and H attention heads collaboratively construct multi-faceted feature representations. For long-term dependency feature modeling, global contextual features are first obtained through pooling operations along the time dimension: Pool ∈ 19 The memory retrieval mechanism is implemented through a trainable memory bank: ∈ 20 where 21 22 The long-term features are then computed as: ℓ softmax √ 23 enabling the model to store and retrieve long-term temporal patterns, thereby enhancing its modeling capability for long- term dependencies. Periodic features are extracted through a dedicated convolutional network: Pool GELU Conv ; 24 The choice of convolution kernel size is based on the theoretical foundations of time series analysis and practical application needs. According to the Nyquist sampling theorem, to capture a signal with frequency ,the sampling frequency must reach 2 which translates to the requirement that the convolution kernel size must satisfy 2 where is the minimum period length. Additionally, considering the effective receptive field: 1 1 25 and the periodic resolution capability 2 / 26 where is the input sequence length. The choice of 2 =7 under the default sequence length = 96 ensures the symmetry of the convolution operation while achieving 0.146 This allows for the capture of the minimum periodic pattern of approximately 7 time steps, while reaching an optimal balance between computational complexity and feature extraction capability. This choice is well-supported by theoretical analysis and experimental validation. Multi-scale features are fused through a dynamic weight mechanism, where 27 and 28 This fusion approach ensures the full utilization of information and stable gradient propagation. The model's computational complexity is , and the spatial complexity is . Through hierarchical feature extraction and adaptive fusion, the model accurately captures the dynamic features of time series across different time scales, providing a new paradigm for time series analysis. Experimental results demonstrate that this multi-scale feature extraction and fusion mechanism achieves significant performance improvements across various time series forecasting tasks. 4. Experiment 4.1. Datasets The southern region of Xinjiang, characterized as a typical temperate continental arid climate zone, exhibits distinctive features including sparse precipitation, intense evaporation, and significant daily temperature variations. The ratio of annual average precipitation to potential evaporation reaches 0.6, indicating that precipitation amounts to approximately 60% of potential evaporation. These unique climatic conditions have resulted in an extremely fragile ecosystem, rendering the region highly sensitive to climate changes and posing significant challenges in ecological conservation. In this study, we selected meteorological data from five representative weather stations: Korla, Kuche, Aral, Kashgar, and Hotan. The dataset was sourced from the meteorological website https://rp5.ru/, spanning from January 2005 to January 2025. Data sampling was conducted at three-hour intervals across eight daily timestamps (02:00, 05:00, 08:00, 11:00, 14:00, 17:00, 20:00, and 23:00), encompassing one temporal feature and 28 meteorological parameters in its dimensional space. Due to multiple factors including sensor malfunctions, network communication issues, and data storage/transmission problems, missing values were inevitable in the raw dataset, particularly prominent in earlier years, though showing significant improvement in recent data collection. Feature elimination was performed for columns with extensive missing data or no data, specifically for parameters including ff10, N, Cl, Nh, Cm, Ch, E, Tg, sss, tR, E, and RRR. For the remaining missing values, considering the smooth nature of weather variations within three-hour sampling intervals, we employed a combination of data imputation techniques including forward fill imputation, random imputation, and linear interpolation to ensure scientific integrity in our data reconstruction process while maintaining the temporal consistency of meteorological measurements. Table 1 presents the definitions and units of the date and meteorological attributes in the dataset and Table 2 details their latitude, longitude, and altitude information. 4.2. Experimental Settings Our model implementation is based on the PyTorch framework, with all experiments conducted on a single NVIDIA GPU for training and evaluation. To ensure experimental reproducibility, we set the random seed to 2021. The specific model hyperparameters are as follows table 3: 39 Table 1. Original data feature information. Contains the meanings of 29 meteorological features and their units, including one date feature and 28 meteorological features. Feature Connotation Unit Date The time at this weather point, taking into account daylight saving time/winter time / T Atmospheric temperature at 2 m above ground level °C Po Atmospheric pressure at the level of the weather station mmHg P Atmospheric pressure at mean sea level mmHg Pa Change in atmospheric pressure during the 3 h preceding the observation mmHg U Relative humidity at 2 m above ground level % DD The wind direction at 10-12 m above ground level during the 10 min prior to observation / Ff Average wind speed at ground level of 10-12 m during the 10 min prior to observation m/s ff10 Maximum gusts at 10-12 meters above ground level during the 10 min prior to observation m/s ff3 Maximum gusts at ground level of 10-12 m between observations m/s N Total cloud cover / WW Weather conditions for the day as announced by the weather station / W1 Past weather during the observation period 1 / W2 Past weather during the observation period 2 / Tn Lowest temperature in the past period (not more than 12 h) °C Tx Maximum temperature in the past period (not more than 12 h) °C Cl Stratocumulus, Stratus, Cumulus and Cumulonimbus / Nh Number of all clouds C1 observed, number of all clouds Cm observed in the absence of cloud C1 / H Cloud base height m Cm Altocumulus, Cirrus and Nimbostratus / Ch Cirrus, Cirrocumulus and Cirrostratus / VV Horizontal visibility Km Td Dew point temperature at 2 m above the ground °C RRR Precipitation within 3 h mm tR Time to reach prescribed precipitation / E Condition of soil surface without snow or ice cover / Tg Minimum temperature of the soil surface at night °C E’ Condition of soil surfaces with snow or ice cover / sss Depth of snow cm Table 2. Information on weather station data Station No. Station Name Latitude (°) Longitude (°) Altitude (m) 51656 Korla 41.46 86.09 944 51644 Kuche 41.43 82.57 1100 51730 Aral 40.33 81.17 1012 57109 Kashgar 39.28 75.59 1269 51828 Hotan 37.07 79.55 1377 Table 3. Model architecture. Parameter Value seq_len 96 pred_len 6, 12, 18, 24, 30, 36, 42, 48 d_model 16 n_heads 4 e_layers 3 d_layers 1 d_ff 32 Moving average window size 25 To comprehensively evaluate the model's predictive performance across various temporal scales, we employed two primary evaluation metrics: Mean Square Error (MSE) and Mean Absolute Error (MAE). Through systematic testing with prediction horizons ranging from 6 to 48 hours, we conducted a thorough assessment of the model's performance in short-term forecasting tasks. This progressive evaluation approach not only demonstrates the model's predictive capabilities in short-term scenarios but also provides substantial evidence for its reliability in real-world applications. Time series forecasting commonly employs three methodological approaches: S (Univariate Forecasting), M (Multivariate Forecasting), and MS (Multivariate-to- Univariate Forecasting), detailed as follows: S (Univariate Forecasting): This approach involves predicting a single variable using only its historical data. The model focuses on the temporal patterns within a single feature's historical observations to forecast future values, based on distributional assumptions. M (Multivariate Forecasting): This method encompasses predicting multiple variables simultaneously using multiple related features. The model leverages historical data from multiple correlated variables to forecast future values for all target variables concurrently. MS(Multivariate-to-UnivariateForecasting):This approach utilizes multiple features to predict a single target variable. Given that precipitation is influenced by multiple factors exhibiting complex non-linear relationships, employing a univariate approach (using only historical precipitation data) would compromise prediction accuracy. Therefore, we opted for the multivariate-to-univariate forecasting method, which incorporates historical weather data from multiple influential factors (such as temperature, humidity, and wind speed) to predict future precipitation levels. By integrating information from multiple features, we can enhance the accuracy of predictions for the single target feature (precipitation). 4.3. Ablation Experiment To conduct ablation studies, we designed the following model variants: •Base: The baseline model 40 •Base_Any:The baseline model enhanced with the AnyVariateAttention module •Base_Any_MTMEM: The Base_Any model augmented with the MTMEM Table 4.The results of the ablation experiments Models Base Base_Any Base_Any_MTMEM Metrics MSE MAE MSE MAE MSE MAE Aral 6h 0.178 0.188 0.175 0.186 0.173 0.184 12h 0.201 0.209 0.198 0.207 0.196 0.205 18h 0.217 0.222 0.214 0.22 0.212 0.218 24h 0.228 0.23 0.225 0.228 0.223 0.226 30h 0.24 0.24 0.237 0.238 0.235 0.236 36h 0.245 0.244 0.242 0.242 0.24 0.24 42h 0.253 0.249 0.25 0.247 0.248 0.245 48h 0.263 0.257 0.26 0.255 0.258 0.253 Kashgar 6h 0.127 0.156 0.124 0.153 0.123 0.152 12h 0.147 0.176 0.144 0.173 0.141 0.171 18h 0.162 0.19 0.159 0.187 0.156 0.184 24h 0.176 0.201 0.173 0.198 0.17 0.195 30h 0.188 0.21 0.185 0.207 0.184 0.205 36h 0.199 0.218 0.196 0.215 0.194 0.214 42h 0.21 0.226 0.207 0.223 0.205 0.222 48h 0.219 0.232 0.216 0.229 0.213 0.225 Korla 6h 0.171 0.185 0.168 0.181 0.167 0.18 12h 0.19 0.206 0.188 0.205 0.186 0.201 18h 0.203 0.218 0.201 0.217 0.2 0.213 24h 0.212 0.226 0.209 0.224 0.207 0.221 30h 0.219 0.233 0.217 0.231 0.215 0.228 36h 0.226 0.24 0.223 0.239 0.221 0.235 42h 0.233 0.245 0.23 0.243 0.226 0.24 48h 0.24 0.252 0.238 0.249 0.235 0.247 Hotan 6h 0.165 0.19 0.169 0.195 0.166 0.193 12h 0.187 0.211 0.189 0.213 0.19 0.215 18h 0.203 0.225 0.207 0.229 0.206 0.227 24h 0.216 0.235 0.219 0.239 0.219 0.238 30h 0.228 0.245 0.231 0.248 0.23 0.246 36h 0.236 0.251 0.241 0.255 0.239 0.254 42h 0.247 0.259 0.248 0.261 0.248 0.26 48h 0.254 0.264 0.254 0.266 0.255 0.267 Kuche 6h 0.15 0.17 0.153 0.175 0.154 0.176 12h 0.166 0.188 0.169 0.192 0.17 0.192 18h 0.178 0.2 0.182 0.205 0.181 0.203 24h 0.188 0.209 0.19 0.212 0.191 0.213 30h 0.196 0.217 0.2 0.222 0.2 0.222 36h 0.201 0.222 0.207 0.228 0.204 0.224 42h 0.207 0.227 0.211 0.232 0.211 0.232 48h 0.212 0.232 0.216 0.236 0.215 0.234 Table 5. Final MSE and MAE Values for Each Region Model Base Base_Any Base_Any_MTMEM Metrics MSE MAE MSE MAE MSE MAE Aral 1.825 1.839 1.801(2.4%) 1.823(1.6%) 1.785(4%) 1.807(3.2%) Kashgar 1.428 1.609 1.404(2.4%) 1.585(2.4%) 1.386(4.2%) 1.568(4.1%) Korla 1.694 1.805 1.674(2%) 1.789(1.6%) 1.657(3.7%) 1.765(4%) Hotan 1.736 1.88 1.758(2.2%) 1.906(2.8%) 1.753(1.7%) 1.9(2%) Kuche 1.498 1.665 1.528(3%) 1.702(3.7%) 1.526(2.8%) 1.696(3.1%) 5. Results and Analysis 5.1. Overall Performance In our comprehensive evaluation of weather data across five representative regions in Xinjiang, we assessed various model variants over prediction horizons ranging from 6 to 48 hours, yielding outstanding results presented in Table 5. The total MSE and MAE values for the three models were calculated and displayed in Table 6. The overall MSE evaluation indicates that the AnyVariateAttention+MTMEM model demonstrates optimal predictive performance, significantly outperforming both the standalone AnyVariateAttention model and the baseline model. Specifically, the performance improvement is most pronounced in the Kashgar region, where the AnyVariateAttention+MTMEM model achieves a 4.2% enhancement compared to the baseline model. This is followed by a 4.0% improvement in the Aral region and a 3.7% increase in Korla. In contrast, both Hotan and Kuche regions 41 exhibited a downward trend, with decreases ranging from 1.7% to 3.7%. In short-term forecasting tasks, the model exhibits exceptional performance advantages, particularly within the 6-hour prediction window, where the AnyVariateAttention +MTMEM model achieves the best performance across all test regions. This advantage is consistently maintained in the 12 to 24-hour prediction range. Notably, even in medium to long-term predictions (30 to 48 hours), although the performance improvement margin slightly narrows, the model retains significant advantages and stable performance. From a regional analysis perspective, Kashgar demonstrates optimal overall performance with the lowest total MSE of 1.386, where the 4.2% improvement brought by MTMEM is particularly significant. The Aral region also shows stable performance improvements and balanced behavior across different time scales. However, in contrast, both Hotan and Kuche regions experienced declines in MSE and MAE. Further analysis revealed that the performance degradation is attributed to data sparsity due to untimely meteorological station records and the impact of climate change, which hindered the historical data's ability to effectively reflect future trends. Additionally, inappropriate model selection, data processing methods, parameter tuning, time lag effects, data noise, and limitations of model assumptions further affected the model's accuracy. These factors collectively hindered the model's ability to accurately capture the complexities of climate change in these two regions, thereby impacting the forecasting results. 5.2. Key Findings Across all regions and prediction horizons, the Multi- Temporal Memory Enhancement Module (MTMEM) consistently delivered performance improvements, ranging from 1.7% to 4.2%. The most significant enhancements were observed in short-term forecasts (6h-24h), with a slower degradation in performance as the prediction horizon increased. The model demonstrated good adaptability across different geographical locations, with the degree of performance improvement varying by region, indicating its ability to capture region-specific weather patterns. These results suggest that the proposed MTMEM module effectively enhances weather forecasting performance, particularly excelling in short-term prediction tasks. The model exhibits stable performance improvements across various geographical locations and time scales, validating its reliability and effectiveness in practical applications. Figures 4 illustrates the visualization of experimental results for the aforementioned five regions, where the blue lines represent the ground truth values and the orange lines indicate the predicted values. Figure 4. Visualization of experiments 6. Conclusion This study presents the TimeMixer-ME model, a dual- memory enhancement framework based on the TimeMixer architecture, which has achieved significant breakthroughs in short-term weather forecasting tasks in the southern Xinjiang region. In prediction tasks using datasets from Aral, Kashgar, and Korla, the model demonstrated substantial performance improvements compared to the baseline TimeMixer, with MSE metrics reduced by 4%, 4.2%, and 3.7%, respectively, and MAE metrics showing comprehensive enhancements ranging from 3.7% to 4.2%. These achievements are attributed to theoretical innovations within the model, particularly the clever integration of the dual-memory mechanism from cognitive science into the time series forecasting domain. By effectively coordinating short-term working memory and long-term reference memory, the model accurately captures the dynamic characteristics of weather systems. Additionally, the designed adaptive feature fusion mechanism allows for flexible adjustment of feature weights based on changes in time scales. However, we also recognize that there is still room for improvement in the model's 42 predictive performance under extreme weather conditions. Future research directions may focus on incorporating external factors such as topography and vegetation to further enhance prediction accuracy, as well as exploring the extension of the model's capabilities to longer time-scale weather forecasting tasks, aiming for a more comprehensive weather prediction capability. Acknowledgments Thanks to Xinjiang University of Science and Technology for its support of the Undergraduate Innovation and Entrepreneurship Program Fund. Project Number: X202513 561108. References [1] Box, G. E. P., Gwilym M. Jenkins, Gregory C. Reinsel and Greta M. Ljung. “Time Series Analysis: Forecasting and Control.” The Statistician 27 (1978): 265-265. [2] Hyndman, Rob J and George Athanasopoulos. “Forecasting: principles and practice.” (2013). [3] Hochreiter, Sepp and Jürgen Schmidhuber. “Long Short-Term Memory.” Neural Computation 9 (1997): 1735-1780. [4] Vaswani, Ashish, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. “Attention is All you Need.” Neural Information Processing Systems (2017). [5] Zhou, Tian, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun and Rong Jin. “FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting.” International Conference on Machine Learning (2022). [6] Zeng, Ailing, Mu-Hwa Chen, L. Zhang and Qiang Xu. “Are Transformers Effective for Time Series Forecasting?” AAAI Conference on Artificial Intelligence (2022). [7] Liu, Yong, Haixu Wu, Jianmin Wang and Mingsheng Long. “Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting.” Neural Information Processing Systems (2022). [8] Wu, Haixu, Teng Hu, Yong Liu, Hang Zhou, Jianmin Wang and Mingsheng Long. “TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis.” ArXiv abs/2210.02186 (2022): n. pag. [9] Shabani, Amin, Amir Hossein Sayyad Abdi, Li Meng and Tristan Sylvain. “Scaleformer: Iterative Multi-scale Refining Transformers for Time Series Forecasting.” ArXiv abs/2206.04038 (2022): n. pag. [10] Lin, Lequan, Zhengkun Li, Ruikun Li, Xuliang Li and Junbin Gao. “Diffusion models for time-series applications: a survey.” Frontiers of Information Technology & Electronic Engineering 25 (2023): 19-41. [11] Liu, Yong, Chenyu Li, Jianmin Wang and Mingsheng Long. “Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors.” ArXiv abs/2305.18803 (2023): n. pag. [12] Lee, Hyunwoo, Seungmin Jin, Hyeshin Chu, Hong Sik Lim and Sungahn Ko. “Learning to Remember Patterns: Pattern Matching Memory Networks for Traffic Forecasting.” ArXiv abs/2110.10380 (2021): n. pag. [13] Liu, Xiangyue, Xinqi Lyu, Xiangchi Zhang, Jianliang Gao and Jiamin Chen. “Memory Augmented Graph Learning Networks for Multivariate Time Series Forecasting.” Proceedings of the 31st ACM International Conference on Information & Knowledge Management (2022): n. pag. [14] Zhang, Yunhao and Junchi Yan. “Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting.” International Conference on Learning Representations (2023). [15] Wang, Shiyu, Haixu Wu, Xiao Long Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y. Zhang and Jun Zhou. “TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting.” ArXiv abs/2405.14616 (2024): n. pag. [16] Xu, Luwen, Jiwei Qin, Dezhi Sun, Yuanyuan Liao and Jiong Zheng. “PFformer: A Time-Series Forecasting Model for Short-Term Precipitation Forecasting.” IEEE Access 12 (2024): 130948-130961. [17] Woo, Gerald, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese and Doyen Sahoo. “Unified Training of Universal Time Series Forecasting Transformers.” ArXiv abs/2402.02592 (2024): n. pag.