# US 2026/0073187 A1 · Time series prediction using convolutional neural network - long short term memory attention model

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0073187 A1
- Document ID: US20260073187A1
- Status: Published application
- Kind: A1
- Application number: 18/882,866
- Filed: 2024-09-12
- Published: 2026-03-12
- Earliest claimed: 2024-09-12
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: Zhao Yu Wang; Peng Hui Jiang; Tian Tian; Mai Zeng; Qian Yi Wu; Si Ling Chen; Wei Dong Zhao; Liang Dong
- Web page: https://patentgrep.com/patent/US20260073187A1

## CPC

- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method for predicting a next time step data element in a set of time series data includes receiving a one-dimensional time series data set and converting the time series data set to a two-dimensional time series data set. The two-dimensional time series data set is provided to an input of a convolutional neural network-long short term memory (CNN-LSTM) model. The CNN-LSTM model generates the next time series step prediction of the two-dimensional time series data. The method compares the prediction to an actual next time series step and responds to a difference exceeding a first threshold by incrementing an outlier counter. When the outlier counter exceeds a predefined size the method alters the CNN-LSTM model. In addition, a visualization of the next time series step prediction of the two-dimensional time series data is generated.

## Claims

### Claim 1

A computer-implemented method comprising:
receiving, at a processor, a one-dimensional time series data set and converting the time series data set to a two-dimensional time series data set using a processor;
providing the two-dimensional time series data set to an input of a convolutional neural network-long short term memory (CNN-LSTM) model using the processor;
generating a next time series step prediction of the two-dimensional time series data using the CNN-LSTM model;
comparing the next time series step prediction to an actual next time series step and responding to a difference between the next time series step prediction and the actual next time series step exceeding a first threshold by incrementing an outlier counter using the processor, and responding to the outlier counter exceeding a predefined count value by altering the CNN-LSTM model; and
generating a visualization of the next time series step prediction of the two-dimensional time series data.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, further comprising normalizing the two-dimensional time series data, providing the normalized two-dimensional time series data to an initial training model and generating a prediction using the initial training model simultaneously with generating the next time series step prediction, comparing an output of the initial training model with the actual next time series step and incrementing the outlier counter using the processor when a difference between the initial training model and the actual time step exceeds a second threshold.

### Claim 3 (depends on 1)

The computer-implemented method of claim 1, wherein the CNN-LSTM comprises:
an input layer for receiving the two-dimensional time series data set and providing the two-dimensional time series data set to a dilated convolutional layer using variable expansion coefficient (VEC-DCNN) configured to generate a dilated convolution of the two-dimensional time series data set; and
a multi-headed self actuation layer receiving the generated dilated convolution and providing multi-headed output to a long short term memory recurrent neural network with host data boundaries (BHI-LSTM) configured to generate the next time series step prediction.

### Claim 4 (depends on 3)

The computer-implemented method of claim 3, wherein the multi-headed self actuation layer includes an embedding layer configured to receive the generated dilated convolution, generate N time series vectors where N is a number of time steps in the generated dilated convolution and to generate a multi-headed output to the BHI-LSTM, wherein the multi-headed output includes N heads with each head corresponding to a distinct component of the N time series vectors.

### Claim 5 (depends on 3)

The computer-implemented method of claim 3, wherein the VEC-CNN includes an input layer for receiving the two-dimensional time series data set and a plurality of sequential dilated convolution layers.

### Claim 6 (depends on 5)

The computer-implemented method of claim 5, wherein the sequential dilated convolution layers have variable dilation rates.

### Claim 7 (depends on 6)

The computer-implemented method of claim 6, wherein altering the CNN-LSTM model includes adjusting a dilation rate of at least convolution layer of the sequential convolution layers.

### Claim 8 (depends on 3)

The computer-implemented method of claim 3, wherein the BHI-LSTM comprises an input layer, a block Hankel conversion layer configured to convert the multi-headed output to a block Hankel tensor, and a long short term memory (LSTM) model with host data boundaries layer including at least two LSTM layers.

### Claim 9 (depends on 8)

The computer-implemented method of claim 8, wherein the block Hankel conversion layer converts the multi-headed output to a block Hankel tensor.

### Claim 10

A method comprising:
receiving a one-dimensional time series data set and converting the time series data set to a two-dimensional time series data set;
providing the two-dimensional time series data set to an input of a convolutional neural network-long short term memory (CNN-LSTM) model;
generating a next time series step prediction of the two-dimensional time series data using the CNN-LSTM model;
comparing the next time series step prediction to an actual next time series step and responding to a difference between the next time series step prediction and the actual next time series step exceeding a first threshold by incrementing an outlier counter, and responding to the outlier counter exceeding a predefined count value by altering the CNN-LSTM model; and
generating a visualization of the next time series step prediction of the two-dimensional time series data.

### Claim 11 (depends on 10)

The method of claim 10, further comprising normalizing the two-dimensional time series data, providing the normalized two-dimensional time series data to an initial training model and generating a prediction using the initial training model simultaneously with generating the next time series step prediction, comparing an output of the initial training model with the actual next time series step and incrementing the outlier counter when a difference between the initial training model and the actual time step exceeds a second threshold.

### Claim 12 (depends on 10)

The method of claim 10, wherein the CNN-LSTM comprises:
an input layer for receiving the two-dimensional time series data set and providing the two-dimensional time series data set to a dilated convolutional layer using variable expansion coefficient (VEC-DCNN) configured to generate a dilated convolution of the two-dimensional time series data set; and
a multi-headed self actuation layer receiving the generated dilated convolution and providing multi-headed output to a long short term memory recurrent neural network with host data boundaries (BHI-LSTM) configured to generate the next time series step prediction.

### Claim 13 (depends on 12)

The method of claim 12, wherein the multi-headed self actuation layer includes an embedding layer configured to receive the generated dilated convolution, generate N time series vectors where N is a number of time steps in the generated dilated convolution and to generate a multi-headed output to the BHI-LSTM, wherein the multi-headed output includes N heads with each head corresponding to a distinct component of the N time series vectors.

### Claim 14 (depends on 12)

The method of claim 12, wherein the VEC-CNN includes an input layer for receiving the two-dimensional time series data set and a plurality of sequential dilated convolution layers.

### Claim 15 (depends on 14)

The method of claim 14, wherein the sequential dilated convolution layers have variable dilation rates.

### Claim 16 (depends on 15)

The method of claim 15, wherein altering the CNN-LSTM model includes adjusting a dilation rate of at least convolution layer of the sequential convolution layers.

### Claim 17 (depends on 12)

The method of claim 12, wherein the BHI-LSTM comprises an input layer, a block Hankel conversion layer configured to convert the multi-headed output to a block Hankel tensor, and a long short term memory (LSTM) model with host data boundaries layer including at least two LSTM layers.

### Claim 18 (depends on 17)

The method of claim 17, wherein the block Hankel conversion layer converts the multi-headed output to a block Hankel tensor.

### Claim 19

A computer program product comprising:
a memory storing instructions for causing a computer system to implement a process including:
receiving a one-dimensional time series data set and converting the time series data set to a two-dimensional time series data set;
providing the two-dimensional time series data set to an input of a convolutional neural network-long short term memory (CNN-LSTM) model;
generating a next time series step prediction of the two-dimensional time series data using the CNN-LSTM model;
comparing the next time series step prediction to an actual next time series step and responding to a difference between the next time series step prediction and the actual next time series step exceeding a first threshold by incrementing an outlier counter, and responding to the outlier counter exceeding a predefined count value by altering the CNN-LSTM model; and
generating a visualization of the next time series step prediction of the two-dimensional time series data.

### Claim 20 (depends on 19)

The computer program product of claim 19, wherein the process further includes normalizing the two-dimensional time series data, providing the normalized two-dimensional time series data to an initial training model and generating a prediction using the initial training model simultaneously with generating the next time series step prediction, comparing an output of the initial training model with the actual next time series step and incrementing the outlier counter when a difference between the initial training model and the actual time step exceeds a second threshold.

### Claim 21 (depends on 19)

The computer program product of claim 19, wherein the CNN-LSTM comprises:
an input layer for receiving the two-dimensional time series data set and providing the two-dimensional time series data set to a dilated convolutional layer using variable expansion coefficient (VEC-DCNN) configured to generate a dilated convolution of the two-dimensional time series data set; and
a multi-headed self actuation layer receiving the generated dilated convolution and providing multi-headed output to a long short term memory recurrent neural network with host data boundaries (BHI-LSTM) configured to generate the next time series step prediction.

### Claim 22 (depends on 21)

The computer program product of claim 21, wherein the multi-headed self actuation layer includes an embedding layer configured to receive the generated dilated convolution, generate N time series vectors where N is a number of time steps in the generated dilated convolution and to generate a multi-headed output to the BHI-LSTM, wherein the multi-headed output includes N heads with each head corresponding to a distinct component of the N time series vectors.

### Claim 23 (depends on 21)

The computer program product of claim 21, wherein the VEC-CNN includes an input layer for receiving the two-dimensional time series data set and a plurality of sequential dilated convolution layers.

### Claim 24

A system comprising:
a client computer having a processor set, a communication fabric and a volatile memory, the volatile memory storing code configured to cause the processor set to generate a time series prediction using a convolutional neural network-long short term memory (CNN-LSTM) attention model by:
receiving, a one-dimensional time series data set and converting the time series data set to a two-dimensional time series data set;
providing the two-dimensional time series data set to an input of the CNN-LSTM model using the processor set;
generating a next time series step prediction of the two-dimensional time series data using the CNN-LSTM model;
comparing the next time series step prediction to an actual next time series step and responding to a difference between the next time series step prediction and the actual next time series step exceeding a first threshold by incrementing an outlier counter, and responding to the outlier counter exceeding a predefined count value by altering the CNN-LSTM model; and
generating a visualization of the next time series step prediction of the two-dimensional time series data.

### Claim 25 (depends on 24)

The system of claim 24, wherein the CNN-LSTM includes an input layer for receiving the two-dimensional time series data set and providing the two-dimensional time series data set to a dilated convolutional layer using variable expansion coefficient (VEC-DCNN) configured to generate a dilated convolution of the two-dimensional time series data set, and a multi-headed self actuation layer receiving the generated dilated convolution and providing multi-headed output to a long short term memory recurrent neural network with host data boundaries (BHI-LSTM) configured to generate the next time series step prediction, and wherein the multi-headed self actuation layer includes an embedding layer configured to receive the generated dilated convolution, generate N time series vectors where N is a number of time steps in the generated dilated convolution and to generate a multi-headed output to the BHI-LSTM, wherein the multi-headed output includes N heads with each head corresponding to a distinct component of the N time series vectors.
