Forecasting timeseries analysis for electrical vehicle battery state utilizing a hybrid transformer architecture
Technologies and techniques for controlling an electrical system via battery forecasting for an electrical system. Time series data is transformed into a series of battery image data including a multi-channel image representing a plurality of battery characteristics and/or battery usage characteristics derived from the time series battery data. An image vector is generated for each respective battery image data of the series, and each image vector is transformed via a first portion of a transformer architecture for sequential data processing and positional encoding. Contextual information is extracted from the transformed image vectors, where each image vector is transformed via a second portion of the transformer architecture using learned weights from the first portion to generate a forecast of future battery characteristics and/or battery usage characteristics. A control command is generated for task-specific processing to modify operation of the electrical system, based on the forecast.
1 . A method for controlling an electrical system via battery forecasting, comprising:
transforming time series battery data into a series of battery image data comprising a multi-channel image representing a plurality of battery characteristics and/or battery usage characteristics derived from the time series battery data;
generating an image vector for each respective battery image data of the series;
transforming each image vector via a first portion of a transformer architecture for sequential data processing and positional encoding;
extracting contextual information from the transformed image vectors;
transforming each image vector via a second portion of the transformer architecture using learned weights from the first portion to generate a forecast of future battery characteristics and/or battery usage characteristics; and
generating a control command for task-specific processing to modify operation of the electrical system, based on the forecast.
2 . The method of claim 1 , wherein the transforming of time series battery data comprises performing grey scale encoding on each of the time series of battery data to obtain the series of battery image data.
3 . The method of claim 2 , wherein generating the image vector comprises the transforming/tokenizing of series of grey scaled encoded battery image data via an autoencoder.
4 . The method of claim 1 , wherein generating the image vector comprises
flattening the battery image data and processing the flattened battery image data via a Convolutional Neural Network (CNN); and
processing the battery image data via an activation function after CNN processing to produce a compressed representation of the battery image data comprising features of interest.
5 . The method of claim 1 , wherein the first portion of the transformer architecture comprises a Masked Embedding Model (MEM) and the second portion of the transformer architecture comprises a Generative Pre-trained Transformer (GPT).
6 . The method of claim 4 , further comprising transferring the learned weights from the MEM to the GPT to generate the forecast of future battery characteristics and/or battery usage characteristics.
7 . The method of claim 1 , wherein the plurality of battery characteristics and/or battery usage characteristics comprise one or more characteristics comprising a state of charge (SoC), state of health (SoH), mileage, temperature, voltage, current, internal resistance, cycle count, charging/discharging rates, and location.
8 . The method of claim 6 , wherein the contextual information comprises information relating to a context representation of target individual embeddings of tokenized subsequences.
9 . The method of claim 1 , further comprising one or more of: (i) processing the extracted contextual information via a classification model to define classes representing different battery states and/or (ii) processing the extracted contextual information via a regression model to extract numerical values or representations from the embeddings, (iii) processing the extracted contextual information via a clustering algorithm to extract clusters characterized by similar battery usage characteristics within each cluster, and (iv) missing data filling of MEM.
10 . An apparatus for controlling an electrical system via battery forecasting, comprising:
a processor;
a communication circuit, operatively coupled to the processor; and
a memory apparatus, operatively coupled to the processor, wherein the processor and memory are configured to
transform time series battery data into a series of battery image data comprising a multi-channel image representing a plurality of battery characteristics and/or battery usage characteristics derived from the time series battery data;
generate an image vector for each respective battery image data of the series;
transform each image vector via a first portion of a transformer architecture for sequential data processing, and positional encoding;
extract contextual information from the transformed image vectors;
transform each image vector via a second portion of the transformer architecture using learned weights from the first portion to generate a forecast of future battery characteristics and/or battery usage characteristics; and
generate a control command for task-specific processing to modify operation of the electrical system, based on the forecast.
11 . The apparatus of claim 9 , wherein the processor and memory are configured to transform time series battery data by performing grey scale encoding on each of the time series of battery data to obtain the series of battery image data.
12 . The apparatus of claim 10 , wherein the processor and memory are configured to generate the image vector by transforming time series battery data that comprises tokenizing the grey scaled encoded battery image data via an autoencoder.
13 . The apparatus of claim 9 , wherein the processor and memory are configured to generate the image vector by
flattening the battery image data and processing the flattened battery image data via a Convolutional Neural Network (CNN); and
processing the battery image data via an activation function after CNN processing to produce a compressed representation of the battery image data comprising features of interest.
14 . The apparatus of claim 9 , wherein the first portion of the transformer architecture comprises a Masked Embedding Model (MEM) and the second portion of the transformer architecture comprises a Generative Pre-trained Transformer (GPT).
15 . The apparatus of claim 13 , wherein the processor and memory are configured to transfer the learned weights from the MEM to the GPT to generate the forecast of future battery characteristics and/or battery usage characteristics.
16 . The apparatus of claim 9 , wherein the plurality of battery characteristics and/or battery usage characteristics comprise one or more characteristics comprising a state of charge (SoC), state of health (SoH), mileage, temperature, voltage, current, internal resistance, cycle count, charging/discharging rates, and location.
17 . The apparatus of claim 15 , wherein the contextual information comprises information relating to a context representation of target individual embeddings of tokenized subsequences.
18 . The apparatus of claim 9 , wherein the processor and memory are configured to preform one or more of: (i) process the extracted contextual information via a classification model to define classes representing different battery states and/or (ii) process the extracted contextual information via a regression model to extract numerical values or representations from the embeddings, (iii) process the extracted contextual information via a clustering algorithm to extract clusters characterized by similar battery usage characteristics within each cluster, and (iv) missing data filling of MEM.
19 . A method for controlling an electrical system via battery forecasting, comprising:
transforming time series battery data into a series of battery image data comprising an encoded multi-channel image representing a plurality of battery characteristics and/or battery usage characteristics derived from the time series battery data;
generating an image vector for each respective battery image data of the series;
transforming each image vector via a Masked Embedding Model (MEM) for sequential data processing and positional encoding;
extracting contextual information from the transformed image vectors;
transforming each image vector via a Generative Pre-trained Transformer (GPT) using learned weights from the first portion to generate a forecast of future battery characteristics and/or battery usage characteristics; and
generating a control command for task-specific processing to modify operation of the electrical system, based on the forecast.
20 . The method of claim 18 , wherein the transforming of time series battery data comprises performing grey scale encoding on each of the time series of battery data to obtain the series of battery image data.