Bank ATM inventory cash prediction method based on deep learning sequence decomposition
By using a deep learning-based sequence decomposition neural network architecture, combined with K-means clustering and Time2Vec cross-attention mechanism, the accuracy problem of ATM cash inventory forecasting was solved, improving forecast accuracy and bank cash management efficiency, and reducing operating costs.
Patent Information
- Application Number
- CN202510996555.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-11-18
AI Technical Summary
Existing technologies struggle to accurately predict ATM cash reserves, leading to either excess or insufficient cash, which impacts bank operating costs and customer service satisfaction. Furthermore, deep learning models face increased training difficulty with large-scale parameters, reducing predictive performance.
A neural network architecture based on deep learning sequence decomposition is adopted. The activity level of ATMs is classified by Kmeans clustering. Combined with Time2Vec and cross-attention mechanism, the time series is decomposed into trend and seasonal components for prediction, and an ATM inventory cash prediction model is constructed.
It significantly improves the accuracy and applicability of ATM cash forecasting, reduces cash redundancy, increases bank cash utilization and customer service satisfaction, and lowers operating costs.
Smart Images
Figure CN120974447A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fully automated products and relates to a method for predicting cash inventory in bank ATMs based on deep learning sequence decomposition. Background Technology
[0002] The booming digital economy is changing consumer shopping habits and diversifying the flow of funds. The emergence of digital currencies and online payments has brought about significant changes to bank cash operations. The demand for cash on hand is influenced by various factors. Allocating funds solely based on subjective forecasts by managers can easily lead to cash surpluses or shortages, wasting human and material resources and hindering efficient capital utilization. Current research rarely addresses the forecasting of ATM cash inventory, and the vast amounts of data stored by banks are not given due attention. The need for accurate forecasting of bank ATM cash inventory and improving capital efficiency urgently requires a solution.
[0003] ATM cash inventory forecasting technology is crucial for bank cash management, directly impacting cash utilization and asset liquidity, and significantly affecting operating costs and profits. This invention studies the daily cash inventory problem in ATMs. When the cash inventory in ATMs is lower than the withdrawal demand for the next day, customer satisfaction will be severely affected, and the bank's reputation and management capabilities will be questioned. Conversely, if excessive cash is locked in ATMs, the bank's cash utilization rate will decrease significantly, the opportunity cost of cash utilization will increase exponentially, bank profits will shrink, and costs for banknote counting, storage, and logistics will also increase.
[0004] In the past, banks relied on subjective estimates of cash on hand by management personnel, which resulted in a large workload, long processing time, and low accuracy. Introducing data-driven time series forecasting models automates the daily cash forecasting task for each ATM, significantly improving the efficiency of bank cash management. However, traditional forecasting models still have significant errors in predicting ATM cash on hand. To ensure service satisfaction, banks still need to maintain a substantial amount of cash reserves.
[0005] Research has revealed that ATM cash inventory data typically exhibits frequent and significant fluctuations, stemming from highly random large-amount cash withdrawals and bank replenishment. This type of high-frequency, high-amplitude, and highly random fluctuation severely impacts the predictive performance of almost all types of models (statistical models, traditional machine learning, and novel deep learning models). Particularly for deep learning models, the large number of parameters can actually increase training difficulty, reducing predictive performance and reliability. Summary of the Invention
[0006] To address the domain-specific challenges of ATM cash forecasting and the lack of in-depth characterization of time series seasonality in traditional ATM cash forecasting methods, as well as other issues, this invention employs the following technical solution: a bank ATM cash forecasting method based on deep learning sequence decomposition, comprising the following steps: The K-means clustering method is used to perform cluster analysis on the time series data of cash balances of ATMs according to different activity levels of ATMs, so as to achieve multiple clusters after the cash balance series data are clustered; A neural network architecture model based on deep learning sequence decomposition is constructed as an ATM inventory cash prediction model. Based on the clustered data, the ATM cash on hand prediction model was trained to obtain a well-trained ATM cash on hand prediction model for judging different levels of ATM activity. The test data is input into a trained ATM cash inventory prediction model used to determine different levels of ATM activity, thereby enabling the prediction of ATM cash inventory.
[0007] Furthermore, it also includes preprocessing analysis of the ATM's cash balance time series data, such as removing duplicates and outliers.
[0008] Furthermore, the process of using the K-means clustering method to perform cluster analysis on the time series data of ATM cash balances according to different activity levels of ATMs is as follows: S21: Calculate the mean and standard deviation of the historical cash inventory sequence of each ATM after preprocessing, and use them as feature vectors; S22: Use the Kmeans clustering method to cluster ATMs based on their feature vectors to obtain a specified number of ATM clusters; S23: For different specified numbers of ATM clusters, visualize the clustering index line chart and the cluster characteristic distribution scatter plot. Based on the clustering index, the number of clusters and the numerical characteristics of the clusters, determine the number of clusters, and realize the clustering analysis of the daily cash balance sequence data of multiple ATMs over a certain period of time according to the different activity levels of ATMs.
[0009] Furthermore, the ATM cash on hand prediction model includes: Standardization module: Used to perform sliding standardization on clustered data using a window-based moving standardization method to alleviate the non-stationarity of the data; Residual connection module: used to process the data input from the standardization module through first-order difference operations to further eliminate data non-stationarity; Time series decomposition module: Used to decompose the data input to the residual connection module into trend components, seasonal components and noise components using time series decomposition technology; Trend prediction module: used to predict the trend components output by the time series decomposition module along the time step dimension using a linear module; The time feature module is used to apply a Time2Vec neural network to the clustered data and output time features. Time-aware cross-attention: Based on the perceptual cross-attention mechanism, seasonal components output by the time series decomposition module and time features output by the time feature module are used to predict seasonal components along the time step dimension. Anti-standardization module: Based on the trend prediction output by the trend prediction module and the seasonal component prediction output by the time-aware cross-attention module, the prediction result is output.
[0010] Furthermore, the process of applying a window-based moving normalization method to the clustered data after clustering, namely sliding normalization, is as follows: For each input sequence of length T, calculate the mean and standard deviation of the current window, and use these to standardize the sequence.
[0011] Furthermore, the process of using time series decomposition technology to decompose the series into trend components and seasonal components is as follows: The trend component is extracted from the sequence input to the residual connection module using the moving average method. The seasonal component is obtained by calculating the difference between the sequence input to the residual connection module and the trend component.
[0012] Furthermore, the perceptual cross-attention mechanism embeds timestamps into a time feature vector using the Time2Vec method, and fuses time features with seasonal component features using the cross-attention mechanism. The input of the perceptual cross-attention mechanism includes query variable Q, keyword variable K, and the value V corresponding to the keyword, and the output is the value O corresponding to query variable Q. When given time window data, the query variable Q is the "time feature" of the next day's date in the time window after being mapped by Time2Vec; The key variable V is the seasonal series obtained by decomposing the ATM balance series within the time window. The variable K represents the "time characteristic" of the date within the time window after being mapped by Time2Vec; Output O is the predicted value of the seasonal sequence portion of the ATM balance for one day within the time window.
[0013] Furthermore, the linear prediction module along the time step dimension predicts the trend components obtained from time series decomposition as follows: By using coefficient regression, the trend components of the time series are modeled to obtain the weights of the linear prediction module; Based on the weights obtained above, the predicted value of the time series trend is obtained through weighted calculation.
[0014] This invention provides a method for predicting ATM cash balances based on deep learning sequence decomposition. Addressing the specific challenges of time-series ATM cash balance prediction, this method significantly improves prediction accuracy and model applicability through an innovative model architecture and time-aware mechanism. This invention employs a neural network architecture based on deep learning sequence decomposition. This type of architecture has strong theoretical basis, high predictive performance, and lightweight structure and parameters. This invention improves parameter utilization through a carefully designed and concise neural network architecture, thereby achieving better predictive performance and interpretability. It has the following advantages: This invention innovatively introduces the Time2Vec neural network to effectively map timestamps into high-dimensional time feature vectors; it also introduces the CrossAttention mechanism to perform feature cross-interaction between the decomposed time series seasonal features and the high-dimensional time features, thereby achieving accurate prediction of seasonal components.
[0015] In addition, this invention employs Residual Connection and sliding normalization to enhance the stationarity of the sequence.
[0016] Through the above-mentioned neural network architecture innovations, this invention improves parameter utilization efficiency and enhances the prediction accuracy of ATM cash reserves while maintaining architectural simplicity and interpretability.
[0017] Based on test data, the model of this invention comprehensively outperforms previous baseline models across multiple general metrics. Furthermore, it has undergone interpretable statistical analysis and experimental verification.
[0018] This method can be deployed in actual banking environments to provide accurate prediction services for ATM cash reserves; it can also serve as a foundational model for future research in related fields, supporting custom feature fusion and architecture extensions. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart of the method; Figure 2 This is a time series graph of the cash on hand at a particular ATM. Figure 3 This is a distribution chart of cash inventory values for the four ATM clusters after clustering. Figure 4 This is a distribution map of cash inventory for each of the four ATM clusters after clustering.
[0021] Figure 5 This is a diagram of the ATM tail box prediction model architecture; Figure 6 This is a schematic diagram of time series decomposition. Detailed Implementation
[0022] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] Figure 1 This is a flowchart of the method; A method for predicting cash reserves in bank ATMs based on deep learning sequence decomposition includes the following steps: S1: Perform desensitization preprocessing on the cash balance sequence data of ATMs; S2: The K-means clustering method is used to perform cluster analysis on the daily cash balance sequence data of multiple ATMs after de-identification and preprocessing over a certain period of time according to the different activity levels of ATMs, so as to achieve multiple clusters after the cash balance sequence data are clustered. S3: Construct a neural network architecture model based on deep learning sequence decomposition as an ATM cash prediction model; S4: Based on the clustered data, train the ATM cash prediction model separately to obtain a well-trained ATM cash prediction model for judging different activity levels of ATMs. S5: Input the test data into the trained ATM cash inventory prediction model used to determine different activity levels of ATMs, thereby predicting the cash inventory of ATMs.
[0025] The steps S1 / S2 / S3 / S4 / S5 are executed sequentially; Furthermore, it also includes preprocessing the daily cash balance sequence data of multiple ATMs over a certain period of time to remove duplicate and outlier values.
[0026] Duplicate values indicate that the balance of a particular ATM appears repeatedly in the data for a specific day or several days. Duplicate data will be removed. Outliers include ATM outages that result in empty or zero balance data for a particular day or several days; such data will be deleted.
[0027] The desensitization preprocessing includes processing the cash balance sequence data of ATMs to hide sensitive information such as XX Bank; Furthermore: The process of using the K-means clustering method to perform cluster analysis on the daily cash balance sequence data of multiple ATMs over a certain period of time according to different activity levels of the ATMs is as follows: S21: Calculate the mean and standard deviation of the historical cash inventory sequence of each ATM after preprocessing, and use them as feature vectors; Given a sequence of ATM cash on hand X ={ x 1, x 2, …, x T}, where T is the total length of the sequence (i.e., the number of historical ATM balance data). x i Indicates the first i Daily ATM cash inventory.
[0028] Calculate: Mean = (x1 + x2 + … + x) T ) / T; The standard deviation is calculated using the following formula:
[0029] Where: N=T is the total length of the sequence; μ: the "mean" obtained from the above formula; σ: standard deviation; S22: The K-means clustering method is used to cluster ATMs based on their feature vectors, resulting in a specified number of ATM clusters. K-means is a commonly used clustering method, first proposed in [year missing], used to group similar samples into one cluster. In this application, an ATM sample point is defined as (mean of ATM cash on hand, standard deviation of ATM cash on hand): K-means groups ATMs with similar means and variances of cash on hand into one cluster, and dissimilar ATMs into different clusters.
[0030] S23: For different specified numbers of ATM clusters, visualize the clustering index line chart and the cluster characteristic distribution scatter plot. Based on the clustering index, the number of clusters and the numerical characteristics of the clusters, determine the number of clusters, and realize the clustering analysis of the daily cash balance sequence data of multiple ATMs over a certain period of time according to the different activity levels of ATMs.
[0031] The clustering metrics include the silhouette coefficient, etc. Numerical characteristics of clusters: mean, standard deviation; For example, when the specified quantity is 2, the activity level of the ATM cluster can be divided into: inactive cluster and active cluster; Active clusters: Clusters with lower means are considered active clusters; Inactive clusters: Clusters with higher mean values are considered inactive. When the specified quantity is 3, the clusters are arranged from highest to lowest according to their mean, corresponding to "inactive," "medium-active," and "active" clusters, respectively. The activity level of ATM clusters can be divided into: Inactive clusters: Clusters with higher mean values are considered inactive. Secondary active clusters: Clusters with means between the higher and lower values are considered secondary active clusters; Active clusters; clusters with lower means are considered active. When the specified quantity is 4, the clusters are arranged from low to high according to their mean, corresponding to "high-frequency active," "medium-active," "low-frequency active," and "inactive" clusters, respectively. The activity level of ATM clusters can be divided into: High-frequency active clusters: Clusters with lower mean values are high-frequency active clusters; Mid-frequency active clusters: those that are larger than the lower mean and smaller than the value between the lower and higher mean are considered mid-frequency active clusters; Low-frequency active clusters: Values larger than the interval between those with lower and higher means, and lower than those with higher means, are considered low-frequency active clusters. Inactive clusters: Clusters with higher mean values are considered inactive. The ATM cash on hand forecasting model includes: Standardization module: Used to perform sliding standardization on clustered data using a window-based moving standardization method to alleviate the non-stationarity of the data; Residual connection module: used to process the data input from the standardization module through first-order difference operations to further eliminate data non-stationarity; The residual connection module, given an input x, and its output y, can be expressed as: y = x + f(x), where f(-) represents a neural network used to predict the difference between y and x. Applied to this application, x is the actual historical ATM balance, y is the predicted ATM balance for the next day, and f(x) is the difference between the current value and the predicted ATM balance for the next day.
[0032] Time series decomposition module: Used to decompose the data input to the residual connection module into trend components, seasonal components and noise components using time series decomposition technology; Trend prediction module: Used to predict the trend components output by the time series decomposition module by using a linear prediction module along the time step dimension. The time feature module is used to apply a Time2Vec neural network to the clustered data and output time features. Time-aware cross-attention: Based on the perceptual cross-attention mechanism, seasonal components output by the time series decomposition module and time features output by the time feature module are used to predict seasonal components along the time step dimension. Anti-standardization module: Based on the trend prediction output by the trend prediction module and the seasonal component prediction output by the time-aware cross-attention module, the prediction result is output.
[0033] Traditional standardization is "global standardization," which assumes that the distribution of all ATM balance time series is within the same range—which is often not applicable in practice.
[0034] Purpose: Sliding normalization considers the distribution of ATM balance time series within a relatively short time window to be within the same range—this is more in line with real-world scenarios. Based on the distribution within this time window, the values of the ATM balance series are mapped to a more stable range, preventing model overfitting, improving the model's generalization ability, and ultimately enhancing the accuracy of subsequent models.
[0035] The process of applying a window-based moving standardization method to the clustered data after clustering, namely the sliding standardization process, is as follows: For each input sequence of length T, calculate the mean and standard deviation of the current window, and use these to standardize the sequence.
[0036] Furthermore, the process of using time series decomposition technology to decompose the series into trend components and seasonal components is as follows: The trend component is extracted from the sequence input to the residual connection module using the moving average (MA) method. Given a sequence X={x1, x2, …, xN} of length N, the moving average method uses a sliding window (of length T) that slides from the leftmost to the rightmost edge of the sequence. For each sequence segment enclosed by the sliding window, its average value is calculated. Specifically, for the window {xt, xt+1, …, xt+T-1}, the mean is obtained as (xt+xt+1+ …+ xt+T-1) / T. The seasonal component is obtained by calculating the difference between the sequence input to the residual connection module and the trend component.
[0037] The perceptual cross-attention mechanism embeds timestamps into time feature vectors using the Time2Vec method, and then fuses time features with seasonal component features using the cross-attention mechanism. The formula for the Time2Vec method is expressed as follows: Given timestamp Time2Vec outputs time features (i.e., one dimension is...) (the vector), the formula is as follows:
[0038] Among them, time features Trainable parameters: , Operator This indicates vector concatenation; Time2Vec is a neural network that maps the time dimension to a vector space using trainable linear variations and trigonometric functions. In this paper, all dates are encoded as vectors, i.e., "time features".
[0039] Cross-attention is also a learnable weighted summation method. The input to the perceptual cross-attention mechanism includes a query variable Q, a keyword variable K, and the value V corresponding to the keyword. The output is the value O corresponding to the query variable Q. When given time window data, the query variable Q is the "time feature" of the next day's date in the time window after being mapped by Time2Vec; The key variable V is the seasonal series obtained by decomposing the ATM balance series within the time window. The variable K represents the "time characteristic" of the date within the time window after being mapped by Time2Vec; Output O is the predicted value of the seasonal sequence portion of the ATM balance for one day within the time window.
[0040] Through a time-aware cross-attention mechanism, the model can query the seasonal component value of the next time step based on the similarity of historical time features.
[0041] The linear prediction module along the time step dimension predicts the trend components obtained from time series decomposition as follows: By using coefficient regression, the trend components of the time series are modeled to obtain the weights of the linear prediction module; Based on the weights obtained above, the predicted value of the time series trend is obtained through weighted calculation.
[0042] Destandardization: the process of restoring the standardized predicted values to the original distribution.
[0043] Figure 2 This is a time series graph of the cash on hand at a particular ATM. Example 1: S1: Acquisition and preprocessing of daily cash balance sequence data; The data in this invention comes from 149 ATMs of a commercial bank in a city in southeastern China. The data collection period is from February 1, 2016 to January 31, 2018, totaling 731 days of daily cash balance data.
[0044] Daily cash balance sequence data is obtained directly from the ATM's sensors, with no missing values or obvious anomalies.
[0045] S2: To improve the efficiency of model training and inference, the KMeans method was used to cluster the cash balance sequences of 149 ATMs into 4 clusters. Data from each cluster were used to train and test the ATM cash inventory prediction model. The clustering features were the mean and standard deviation of the sequences to reflect the distributional similarity between different sequences.
[0046] Figure 3 This is a distribution chart of cash inventory values for the four ATM clusters after clustering. Figure 4 This is a distribution map of cash on hand for each of the four ATM clusters after clustering, including the main distribution range of ATM balances, i.e., the boxes and box lines; the mean of ATM balances, i.e., the yellow line; and the outliers of ATM balances, i.e., the small dots above the box lines. Figure 5 This is a diagram of the ATM tail box prediction model architecture; Figure 6 This is a schematic diagram of time series decomposition; S3: Construct a neural network architecture model based on deep learning sequence decomposition as an ATM cash prediction model; The RTDLinear model (a linear prediction model based on residual connectivity, time awareness, and sequence decomposition enhancement) consists of the following five modules: (1) Standardization module and anti-standardization module To mitigate the non-stationarity of the time series, the model employs a window-based moving standardization method. Specifically, for each input sequence of length T, the mean and standard deviation of the current window are calculated, and the sequence is standardized accordingly.
[0047] The destandardization process restores the standardized predicted values to the original distribution.
[0048] The standardized formula: xi is the ATM cash inventory value on day i; μ: The average value of the ATM cash inventory sequence within the current time window; σ: Standard deviation of the ATM cash inventory sequence within the current time window; (2) Residual connection module The residual connection module aims to further eliminate the impact of non-stationarity on the model. This module receives a standardized input sequence and processes it through a first-order differencing operation. The output of the residual connection module is used as the input for subsequent modules.
[0049] (3) Time series decomposition module This invention employs time series decomposition technology to break down the series into trend components, seasonal components, and noise components. The trend component is extracted using a moving average (MA) operation. Subsequently, the seasonal component is obtained by calculating the difference between the original series and the trend component.
[0050] (4) Trend Prediction Module For the prediction of trend components, a linear module is used to make predictions along the time step dimension; (5) Time Feature Module: Used to apply the Time2Vec neural network to the clustered data and output time features; (6) Time-perception cross attention To better capture the temporal information of seasonal components, this invention proposes a time-aware cross-attention mechanism. This mechanism embeds timestamps into time feature vectors using the Time2Vec method and fuses time features with seasonal component features using a cross-attention mechanism.
[0051] Through a time-aware cross-attention mechanism, the model can query the seasonal component value of the next time step based on the similarity of historical time features.
[0052] S4. Based on the clustered data, train the ATM cash prediction model to obtain a well-trained ATM cash prediction model for judging different levels of ATM activity. The RTDLinear model is implemented using the PyTorch framework, with mean squared error (MSE) used as the loss function during training. The training dataset comprises 70% of the data in each cluster, the validation set 10%, and the test set 20%. The input sequence length is 24, the temporal embedding dimension is 5, and the scaling factor in the cross-attention mechanism is 0.1. During training, the model is trained for a maximum of 800 epochs to ensure a stable decrease in the loss curve, and the model with the lowest loss on the validation set is selected as the final model.
[0053] S5. Implementation Results Benchmarking on test sets across four clusters, the RTDLinear model significantly outperforms other baseline models, including DLinear, GRU, LSTM, MLP, SVR, and ARIMA, across all evaluation metrics, including scaled root mean square error (S-RMSE), scaled mean absolute error (S-MAE), and symmetric mean absolute percentage error (SMAPE). Furthermore, ablation experiments validate the performance enhancements provided by the time-aware cross-attention and residual connection modules. Experimental results demonstrate that the RTDLinear model effectively addresses the specific challenges of ATM cash balance time series data, offering an efficient and accurate solution for ATM cash balance prediction.
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting cash reserves in bank ATMs based on deep learning sequence decomposition, characterized in that: Includes the following steps: The K-means clustering method was used to perform cluster analysis on the time series data of ATM cash balances according to different activity levels of ATMs, resulting in multiple clusters after clustering the ATM cash balance data; Construct a deep learning-based neural network architecture model as an ATM inventory cash prediction model; Based on the clustered data, the ATM cash on hand prediction model was trained to obtain a trained ATM cash on hand prediction model for judging different levels of ATM activity. The test data is input into a trained ATM cash inventory prediction model used to determine different levels of ATM activity, thereby enabling the prediction of ATM cash inventory.
2. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 1, characterized in that: It also includes preprocessing analysis of the time series data of the cash balance of the ATM to remove duplicate and outlier values.
3. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 1, characterized in that: The process of using the K-means clustering method to perform cluster analysis on the time series data of ATM cash balances according to different activity levels of ATMs is as follows: S21: Calculate the mean and standard deviation of the historical tail box balance time series data of each ATM after preprocessing and analysis, and use them as feature vectors; S22: Use the Kmeans clustering method to cluster ATMs based on their feature vectors to obtain a specified number of ATM clusters; S23: For different specified numbers of ATM clusters, visualize the clustering index line chart and the cluster characteristic distribution scatter plot. Based on the clustering index, the number of clusters and the numerical characteristics of the clusters, determine the number of clusters, and realize cluster analysis of the cash balance time series data of ATMs according to different activity levels of ATMs.
4. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 1, characterized in that: The ATM cash on hand prediction model includes: Standardization module: Used to perform window-based moving standardization on clustered data after clustering, mitigating the non-stationarity of the data; Residual connection module: used to process the data input from the standardization module through first-order difference operations to further eliminate data non-stationarity; Time series decomposition module: Used to decompose the data input to the residual connection module into trend components, seasonal components and noise components using time series decomposition technology; Trend prediction module: used to predict the trend components output by the time series decomposition module along the time step dimension using a linear module; The time feature module is used to apply a Time2Vec neural network to the clustered data and output time features. Time-aware cross-attention: Based on the perceptual cross-attention mechanism, seasonal components output by the time series decomposition module and time features output by the time feature module are used to predict seasonal components along the time step dimension. Anti-standardization module: Based on the trend prediction output by the trend prediction module and the seasonal component prediction output by the time-aware cross-attention module, the prediction result is output.
5. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 4, characterized in that: The process of applying a window-based moving standardization method to the clustered data after clustering, namely the sliding standardization process, is as follows: For each input time series data, the time window length is set according to week or month, the mean and standard deviation of the current window are calculated, and the series is standardized accordingly.
6. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 4, characterized in that: The process of using time series decomposition technology to decompose the series into trend components and seasonal components is as follows: The trend component is extracted from the sequence input to the residual connection module using the moving average method. The seasonal component is obtained by calculating the difference between the sequence input to the residual connection module and the trend component.
7. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 4, characterized in that: The perceptual cross-attention mechanism embeds timestamps into a time feature vector using the Time2Vec method, and then fuses the time features with seasonal component features using the cross-attention mechanism. The input to the perceptual cross-attention mechanism includes a query variable Q, a keyword variable K, and the value V corresponding to the keyword; the output is the value O corresponding to the query variable Q. When given time window data, the query variable Q is the "time feature" of the next day's date in the time window after being mapped by Time2Vec; The key variable V is the seasonal series obtained by decomposing the ATM balance series within the time window. The variable K represents the "time characteristic" of the date within the time window after being mapped by Time2Vec; Output O is the predicted value of the seasonal sequence portion of the ATM balance for one day within the time window.
8. The method for predicting bank ATM cash reserves based on deep learning sequence decomposition according to claim 1, characterized in that: The linear prediction module along the time step dimension predicts the trend components obtained from time series decomposition as follows: By using coefficient regression, the trend components of the time series are modeled to obtain the weights of the linear prediction module; Based on the weights obtained above, the predicted value of the time series trend is obtained through weighted calculation.