Multi-Task Deep Learning for Short-Term Load Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accurate short-term electric load forecasting is challenging due to the intermittency of renewable energy sources and increasing uncertainties from electric vehicle charging and high-demand appliances, which affects the efficiency and cost of power grid operations.
Innovation Solution
A method and apparatus using multi-task deep learning that clusters commodity-consuming objects based on environmental and calendar data, employing a Long Short-Term Memory (LSTM) based multi-task learning model to predict future energy consumption by training on historical data, thereby improving forecasting accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional single-task learning models are used for load forecasting, then the model structure is simple and easy to implement, but the forecasting accuracy deteriorates due to inability to capture multiple influencing factors simultaneously
Solution Approach 1:
The patent segments the load forecasting problem into multiple independent tasks (individual load prediction, aggregate load prediction, renewable energy prediction, EV charging prediction) and assigns dedicated neural network branches to each task. This segmentation allows each sub-model to specialize in capturing specific patterns while the overall multi-task framework integrates them, resolving the contradiction between model complexity and forecasting accuracy by organizing complexity in a structured, manageable way.
Solution Approach 2:
The patent implements a universal multi-task learning framework where a shared base model learns common temporal patterns and features that are useful across all forecasting tasks, while task-specific branches adapt to individual requirements. This multi-functionality allows a single model to perform multiple forecasting functions simultaneously, improving accuracy without proportionally increasing overall complexity.
2Measurement precision
If more data features and tasks are incorporated into the forecasting model, then the forecasting accuracy improves, but the computational cost and training time increase
Solution Approach 1:
The patent merges multiple forecasting tasks into a unified multi-task learning model with shared computational layers. By combining individual load forecasting, aggregate load forecasting, renewable energy forecasting, and EV charging forecasting into a single integrated model, the system avoids redundant computations that would occur if separate models were trained independently, thus reducing overall computational cost while maintaining high accuracy across all tasks.
Solution Approach 2:
The patent incorporates preliminary feature engineering and data preprocessing steps that extract and organize relevant features before they enter the neural network. By pre-processing data to identify key temporal patterns, seasonal trends, and correlations upfront, the model requires fewer computational resources during training and inference, reducing the energy cost while preserving forecasting accuracy.
3Adaptability or versatility
If the model is trained on cluster-specific data only, then the training data is simple to manage, but the model's adaptability to different clusters deteriorates
Solution Approach 1:
The patent changes the parameter of data representation by introducing cluster identifiers and cluster-specific feature transformations. Instead of managing completely separate datasets for each cluster, the model uses parameter changes to adapt the same base model to different clusters through learned cluster-specific weights and biases, achieving high adaptability while keeping data management relatively simple.
Solution Approach 2:
The patent creates copied versions of the base model for each cluster, where each copy is initialized with cluster-specific data but shares the same architectural structure. This copying approach allows the model to be trained on cluster-specific patterns while maintaining consistency across clusters, achieving adaptability without requiring entirely different models for each cluster.
Data Source
AI summary
A method of load forecasting using multi-task deep learning includes obtaining references data corresponding to commodity consuming objects, clustering the commodity consuming objects into clusters based on the obtained reference commodity consumption data; obtaining cluster models based on: reference commodity consumption data, reference environmental data, and reference calendar data; inputting, into the cluster models, present data corresponding to the commodity consuming objects; and predicting, based on an output of the cluster models, a future commodity consumption for the commodity consuming objects. The cluster models include multi-task learning processes having joint loss functions.