💻 computer science

Inpainting-Style Conditional Diffusion for Multivariable Time Series Forecasting

This paper proposes a novel conditional diffusion framework that reformulates multivariable solar power forecasting as an image inpainting problem by converting temporal data into 2D patches, enabling a U-Net-based model to accurately reconstruct future time steps from historical observations through a mask-guided denoising process.

Original authors: Kourosh Kiani, S. M. Muyeen

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kourosh Kiani, S. M. Muyeen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict exactly how much sunlight a solar panel will catch tomorrow. It's a bit like trying to guess the plot of a movie you haven't seen yet, but you only have the first half of the script. The problem is that the "weather movie" is chaotic; clouds move fast, the sun hides, and the energy output jumps around unpredictably.

This paper introduces a new way to solve this puzzle using a clever trick borrowed from art restoration and a powerful type of artificial intelligence called a Diffusion Model.

Here is the breakdown of their approach in simple terms:

1. The Problem: The "Blurry Future"

Traditional methods for predicting solar power are like trying to draw a picture of a moving car by looking at a single, blurry snapshot. They often struggle because solar energy is messy, non-linear, and changes rapidly.

2. The Solution: Turning Data into a Picture

The authors realized that standard computer programs for images (specifically a type called U-Net) are very good at seeing patterns. So, they decided to stop treating solar data as a long list of numbers and instead turn it into a 2D image.

The Analogy: Imagine taking a strip of graph paper showing 16 different weather variables (like temperature, wind, and sun) over time. They stack these strips on top of each other to create a square grid, like a small pixelated photo.
The Trick: To make this "photo" fit perfectly into the computer's brain (the U-Net), they add a border of "invisible zeros" (padding) around the edges, just like adding a white border to a photograph so it fits in a specific frame.

3. The Core Idea: "Inpainting" (The Art Restoration Metaphor)

This is the most creative part of the paper. Instead of asking the AI, "What comes next?" (which is hard), they ask it, "Can you fix this torn picture?"

The Setup: They take their "solar photo" and cover the bottom half (the future) with a mask, like a piece of black tape. The top half (the past/historical data) is left visible.
The Process (Diffusion):
1. The Mess: They take the hidden future part and slowly add "digital noise" to it until it looks like static on an old TV screen.
2. The Training: They teach the AI to look at this noisy, half-covered picture and guess what the "static" looks like.
3. The Cleanup (Reverse Diffusion): When it's time to predict, they start with a blank, noisy canvas for the future. The AI looks at the clear, visible past (the top half) and the noisy future, then slowly "dusts off" the noise, step-by-step, until the future picture is clear and makes sense based on the past.

Think of it like a detective looking at a crime scene photo where the future is blurred out. The detective uses the clear clues from the past to slowly sharpen the blurry part until the whole story is visible.

4. Why This Works Better

The paper tested this method on real solar data from Australia (the GEFCom2014 dataset) and compared it to many other "smart" methods, including standard math formulas, older machine learning, and even the newest "Transformer" models.

The Result: Their "Inpainting Diffusion" method was the most accurate.
Why?
- It sees the whole picture: By turning data into a 2D image, the AI can see how different weather variables interact with each other (like how wind and clouds work together) better than methods that just look at one line of numbers.
- It handles uncertainty: Because it uses a "generative" process (creating from noise), it naturally understands that the future isn't just one single number, but a range of possibilities, making it more robust when the weather is weird.
- It uses all the data: They showed that using all 25 available weather features (instead of just 16) and padding the image correctly made the predictions even sharper.

5. The Catch

Just like any prediction, the further into the future you look, the harder it is. The paper notes that the model is excellent at predicting the next hour or two (short-term), but as you try to predict further out (8 hours ahead), the accuracy naturally drops a little, just like trying to guess the ending of a movie gets harder the more plot twists you have to imagine.

Summary

In short, the authors built a system that treats solar power forecasting like restoring a damaged painting. They take the known past, cover the unknown future, turn the whole thing into a grid, and use a special AI to "clean the noise" off the future part, guided by the clear past. This approach proved to be more accurate than almost every other method currently used for predicting solar energy.

Technical Summary: Inpainting-Style Conditional Diffusion for Multivariable Time Series Forecasting

Problem Statement
The integration of high-penetration photovoltaic (PV) systems into modern power grids introduces significant operational challenges due to the inherent intermittency, nonlinearity, and stochastic nature of solar energy. While accurate forecasting is critical for grid stability and energy management, conventional approaches face limitations. Statistical methods (e.g., ARIMA) struggle with nonlinear dynamics; physics-based models rely on complex preprocessing and suffer from uncertainties in meteorological inputs; and standard deep learning models (e.g., LSTM, GRU, Transformers) often rely on deterministic predictions that may fail to adequately capture the multi-modal distributions and uncertainty inherent in renewable generation. Furthermore, existing data-driven methods often require extensive handcrafted feature engineering or struggle to model complex spatiotemporal dependencies efficiently.

Methodology
The paper proposes a novel framework that reformulates multivariable solar power forecasting as a time-series inpainting problem using Conditional Denoising Diffusion Probabilistic Models (DDPM). The core methodology involves three primary components:

Structured Spatiotemporal Representation:
Instead of processing raw one-dimensional time-series data, the authors transform multivariate PV data into structured two-dimensional image representations. Using a sliding-window approach, a sequence of time steps and features is arranged into a patch (e.g., $16 \times 16$ or $32 \times 32$ ). To ensure compatibility with the U-Net architecture's requirement for dimensions divisible by powers of two, a zero-padding strategy is introduced. This allows the model to utilize all available features (e.g., all 25 features from the GEFCom2014 dataset) without discarding information, while maintaining the fixed input size required for hierarchical encoding and decoding.
Mask-Based Conditional Diffusion (Inpainting Formulation):
The forecasting task is treated as an image inpainting problem. The input patch is partitioned into two regions:
- Conditioning Region (Fixed): Historical observations (upper rows) are preserved and remain uncorrupted throughout the diffusion process.
- Target Region (Noised): Future time steps (lower rows) are progressively corrupted with Gaussian noise during the forward diffusion process.
During training, the model learns to predict the injected noise given the noisy target region and the fixed historical context. During inference (reverse diffusion), the model starts with pure noise in the target region and iteratively denoises it, conditioned on the preserved historical data, to reconstruct coherent future sequences. This approach replaces deterministic regression with a probabilistic generative process.
Training and Architecture:
The framework employs a U-Net architecture trained with a supervised denoising objective (Mean Squared Error between true and predicted noise). The loss function is applied only to the target region, ensuring the model focuses on reconstructing future values while respecting the historical context.

Key Contributions

Inpainting-Based Forecasting: The paper introduces a paradigm shift from direct sequence-to-sequence prediction to a conditional inpainting formulation. By treating future values as missing regions to be reconstructed, the model leverages the robustness of diffusion models to generate context-consistent trajectories.
Padding-Aware Spatiotemporal Representation: A novel data representation strategy is proposed that converts multivariate time-series data into fixed-size 2D patches using zero-padding. This design preserves all original features and enables the application of U-Net-based diffusion models, which are typically designed for image data, to time-series forecasting tasks.
Integration of Conditional Diffusion: The work demonstrates the effective application of conditional DDPMs for renewable energy forecasting, showing how mask-guided diffusion can capture complex temporal dynamics and inter-feature dependencies better than traditional deterministic models.

Experimental Results
The proposed framework was evaluated on the GEFCom2014 benchmark dataset (Zone 1, Australia), which includes hourly PV generation and meteorological variables.

Performance Metrics: The model was assessed using Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and Pearson correlation.
Ablation Studies: Experiments confirmed that utilizing the full set of 25 features with symmetric zero-padding (creating a $32 \times 32$ input) yielded superior performance compared to using only 16 features or asymmetric padding. Short-term forecasting horizons (1–2 time steps) achieved the highest accuracy, with a Pearson correlation of 0.994, MAPE of 0.017, and RMSE of 0.022.
Comparative Analysis: The proposed method outperformed a wide range of state-of-the-art baselines, including statistical models (ARIMA), machine learning algorithms (SVR), deep learning architectures (LSTM, GRU, CNN-LSTM), and advanced Transformer variants (Informer, Autoformer, FEDformer). Notably, it also surpassed an unconditional diffusion model, highlighting the importance of the conditioning mechanism.
Sensitivity Analysis: Results indicated that forecasting accuracy degrades gracefully as the prediction horizon increases, a behavior attributed to the accumulation of uncertainty in the iterative reverse diffusion process. The analysis also confirmed that sufficient historical context (more conditional rows) is critical for high-fidelity reconstruction.

Significance and Claims
The paper claims that integrating diffusion-based generative modeling with an inpainting formulation offers a robust, flexible, and high-fidelity alternative for solar power forecasting. By moving away from deterministic regression, the proposed method better captures the stochastic nature of solar generation and the uncertainty inherent in future states. The authors assert that their structured spatiotemporal representation allows for the effective learning of both temporal dynamics and inter-feature dependencies, which are often underutilized in conventional models. While the method demonstrates superior accuracy, particularly for short-term horizons, the authors modestly note that future work could focus on incorporating multi-modal inputs (e.g., satellite imagery), integrating attention mechanisms for long-range dependencies, and optimizing computational efficiency for real-time deployment.