We just created LSTEnergy [last energy], a Long-Term-Short-Memory energy forecasting model built using Blossom Sky and real-world smart meter data acquired in 2020. LSTEnergy can be found on GitHub [1] and on our HuggingFace space [2].

Why did we develop an industry-specific model?

LLM and LSTM-based models are a key component of the current AI craze, thanks to OpenAI and ChatGPT. Companies are beginning to establish AI strategies, sometimes without knowing what they should do or how to achieve it. Using large-scale models, which are strong but not specialized, frequently leads to greater confusion since they are language-based models that just give discussions. Conversations are no longer very useful when a corporation wishes to optimize particular processes with AI.

With LSTEnergy, we hope to help businesses better understand their energy usage, estimate it more accurately, and therefore save energy and CO2. LSTEnergy is a time-series forecasting model that predicts future events using time-based previous data. Using Blossom Sky, this model may forecast likely future consumption by running on numerous, independent data stores or data lakes in a sliding time frame method.

Depending on the dataset, LSTEnergy executes with high probability after around 20 epochs. The model is typically performed once every week per smart meter. To utilize LSTM for time series forecasting, we must first train it on a collection of past data and then attempt to predict future values. Using the ‘smartmeter’ dataset, which is part of our open source model and available in our GitHub as well as HuggingFace space, the LSTEnergy model trains a function that transfers a series of prior observations as input to an output observation.

The ‘smartmeter’ data set is a collection of energy usage data that has been collected over a longer period of time. LSTEnergy is trained to learn a function that transfers a series of previous consumption in x = days as input to an output of predictive energy consumption. For example, we may take the previous ten days’ consumption as input and forecast the eleventh day’s consumption as output.

Why did we adopt an LSTM approach?

Long Short-Term Memory (LSTM) is a form of recurrent neural network (RNN) that can learn long-term dependencies. LSTM models have a unique design in that they employ memory cells and gates to control the flow of information. This enables people to recall vital historical knowledge while discarding unnecessary information. They are particularly useful for time series forecasting, in which the aim is to predict future values based on previous events. In the context of time series forecasting, an LSTM model takes a sequence of previous observations as input and predicts the next value in the sequence as output. The model is trained on historical data to discover underlying patterns and correlations between input characteristics and the target variable. After successfully training the model, it may be used to make predictions based on fresh data sets without having to be trained again. A user may adjust the number of layers or the number of neurons per layer to increase the accuracy of LSTEnergy.

LSTM is now a member of the neural network family. However, RRNs have a tendency to forget information that is too distant in the past. This is due to the fact that the hidden state vector is diluted by repeated multiplications and additions as it traverses the network. This is known as the “vanishing gradient” problem, because it inhibits RNNs’ capacity to learn long-term relationships.

This issue is addressed by LSTM by introducing a new component: a cell state vector. The cell state functions as a memory, storing and retrieving information over lengthy periods of time. These gates are neural networks that learn to regulate whether information from the cell state and the hidden state to preserve or reject.

About Databloom

Databloom is a software company that has developed a powerful AI-Powered Distributed Data Processing Platform Integration as a Service, Blossom Sky [3]. Blossom Sky enables users to unlock the full potential of their data by connecting data sources, enabling AI, and gaining performance by running data processing and AI directly at independent data sources. Blossom Sky allows for data collaboration, increased efficiency, and new insights by breaking data silos in a unified manner through a single system view. The platform is designed to adapt to a wide variety of AI algorithms and models.‍

[1] databloom-ai/LLM-LSTM: Specialized LLM / LSTM models

[2] databloom/LSTEnergy · Hugging Face

[3] Federated Data Processing Platform | Blossom Sky (databloom.ai)