# MTMF-Grid: A multi-task multi-modal fusion model for operational forecasting and decision support in power grids

## TL;DR

The paper introduces MTMF-Grid, a model for power grid investment decisions using multi-task and multi-modal data fusion.

## Contribution

A novel multi-task multi-modal fusion model (MTMF-Grid) is proposed for operational forecasting and decision support in power grids.

## Key findings

- MTMF-Grid outperforms baseline models on SEWA and OPSD datasets.
- Achieves 3.06% MAPE for electricity price prediction and 0.915 accuracy for load fluctuation risk classification.

## Abstract

Power grid strategic emerging business investment features multi-objective coupling and multi-source heterogeneous data. It requires simultaneous completion of regression and classification tasks, making traditional single-task or single-modal assessment methods inadequate for precise decision-making. This study proposes a multi-task multi-modal fusion model (MTMF-Grid) for operational forecasting and decision support in strategic emerging power grid investments. MTMF-Grid leverages operational data proxies to support investment decision-making, rather than directly predicting financial returns. MTMF-Grid adopts a modular architecture with three core mechanisms: a task-adaptive Transformer to balance general and task-specific feature expression, a Cross-Fusion Gating Mechanism (CFGM) for dynamic multi-modal fusion and robustness to modal missing scenarios, and a loss variance-based mechanism to dynamically adjust task weights and alleviate gradient conflicts. Experiments on SEWA (Sharjah Electricity and Water Authority dataset) and OPSD (Open Power System Data) datasets show MTMF-Grid outperforms mainstream baseline models. It achieves 3.06% MAPE (Mean Absolute Percentage Error) for hourly electricity price prediction, 0.915 accuracy for load fluctuation risk classification. This study presents a comprehensive framework for supporting strategic decision-making in power grid investment through operational forecasting and multi-modal data integration.