Localized Weather Variability in Forecasting

Thrilled to unveil our latest work: multi-modal machine learning to forecast localized weather! We construct a graph neural network to learn dynamics at point locations, where typical gridded forecasts miss significant variation. Paper: https://lnkd.in/eBmfsJin Weather dataset: https://lnkd.in/ejCG8bKs Code: https://lnkd.in/eQg-JzQJ AI weather models have made huge strides, but most still emulate products like ERA5, which struggle to capture near-surface wind dynamics. The correlation between ERA5 and ground weather station data is low due to topography, buildings, vegetation, and other local factors. In this work, we forecast near-surface wind at localized off-grid locations using a message-passing graph neural network ("MPNN"). The graph is heterogeneous, integrating both global forecasts (ERA5) and historical local weather station data as different nodes. What do we find? First off, ERA5 interpolation performs poorly, failing to capture local wind variations, especially in coastal and inland regions with complex conditions. An MLP trained on historical data at a location performs better than ERA5 interpolation, as it learns from the station's past observations. However, it struggles with longer lead times and lacks the spatial context necessary to capture weather patterns. Meanwhile, our MPNN dramatically improves performance, reducing the error by over 50% compared to the MLP. This is because the MPNN incorporates spatial information through message passing, allowing it to learn local weather dynamics from both station data and global forecasts. Interestingly, adding ERA5 data to the MLP does not improve its performance significantly. The MLP struggles to integrate spatial information from global forecasts, while the MPNN excels, highlighting the importance of combining global and local data. Large improvements in forecast accuracy occur at both coastal and inland locations. Our model shows a 92% reduction in MSE relative to ERA5 interpolation overall. This work showcases the strength of machine learning in combining multi-modal data. By using a graph to integrate global and local weather data, we were able to generate much more accurate localized weather forecasts! Congrats to Qidong Yang and Jonathan Giezendanner for the great work, and thanks to Campbell Watson, Daniel Salles Chevitarese, Johannes Jakubik, Eric Schmitt, Anirban C., Jeremy Vila, Detlef Hohl, and Chris Hill for a wonderful collaboration. Thanks also to our partners at Amazon Web Services (AWS) for providing cloud computing and technical support!

Bridging the gap between global weather models and local reality Accurate weather forecasts at specific locations are critical for wildfire management, renewable energy, agriculture, and infrastructure planning. Yet most numerical and AI-based weather models still operate on coarse grids, systematically missing near-surface, local effects — especially for wind. A recent study, “Local Off-Grid Weather Forecasting With Multi-Modal Earth Observation Data”, demonstrates a powerful alternative: instead of simply downscaling gridded forecasts, the authors correct large-scale weather predictions using direct station-level observations. By combining: • historical measurements from weather stations • large-scale numerical forecasts (ERA5 / HRRR) • and a transformer model with dynamic spatial attention the approach delivers highly accurate off-grid forecasts at irregular station locations. The results are striking — up to 80% error reduction for near-surface wind compared to gridded forecasts alone. The key takeaway: Even the best global or ML weather models cannot achieve local accuracy without direct station inputs. Transformers excel here because they can dynamically learn which nearby observations matter most under changing conditions. This work points toward a future where local weather intelligence is no longer limited by grid resolution — enabling better decisions in high-stakes, location-sensitive applications. Source: https://lnkd.in/dp5bSKWa #WeatherForecasting #EarthObservation #AI #Transformers #ClimateTech #RenewableEnergy #WildfireManagement #MachineLearning

Local Weather Data x Critical Risk Management We talk a lot about environmental impacts on high-risk activities—like wind speed & direction impacting crane lifts, work at height, and heavy equipment operations—but how representative is the weather data we rely on? Most of the time, we use forecasted conditions from national meteorological services which are great for general awareness but often don’t reflect site-specific conditions. A forecast from a weather station 30km away doesn’t capture sudden wind gusts at a crane lift zone, temperature variations on-site, or microclimates created by terrain. Having local, real-time weather data at the actual worksite enables better risk management decisions. Instead of relying on broad forecasts, organisations can monitor live conditions at the precise location where critical work is happening. PLUS you get your own comprehensive data set for analytics... In the photos I'm holding a Davis EnviroMonitor Gateway LTE & Vantage Pro2 GroWeather Sensor Suite which is an example of a local weather monitoring system. This system provides real-time, hyper-local weather data directly from the worksite, enabling data-driven risk management decisions. It delivers real-time updates every 2.5 seconds; has wind speed, temperature, humidity, and rainfall monitoring plus solar radiation and evapotranspiration data which is also valuable for heat stress risk. This model has LTE connectivity (basically you can stick a SIM card in it) for remote monitoring and integration with cloud platforms. These systems aren't that expensive and offer new insights for local risk management that I've found can make a pretty big difference to your risk control strategy. Is anyone else implementing local weather systems for crane ops or other critical risk management? #safetytech #safetyinnovation #IoT

Although nearly a month has passed since the heavy precipitation event that struck the #Valencia region, I believe it’s still relevant to analyze some key aspects of this occurrence. This time, rather than examining the historical data from nearby stations (https://lnkd.in/emavXauX), I want to focus on the spatial distribution of accumulated precipitation, particularly in the context of predictability. Obtaining official data from AEMET is a challenge, but I’ve compiled observations from various weather networks for this analysis. In the plot, each point represents a station and is color-coded to indicate the total precipitation recorded on October 29th, with colors ranging from white to yellow, blue, and red. The area with the highest precipitation was centered between #Turis and #Chiva, where values reached up to 640 mm. The official AEMET station in Turis even recorded an astounding 771 mm. In stark contrast, Valencia itself only recorded about 10 mm—an extraordinary difference given the mere 20 km separating these locations. Now you may ask yourself what are those triangles in the background. The smaller triangles represent the grid of ICON-EU, one of Europe’s state-of-the-art weather forecasting models. Since this model assigns only a single value to each grid cell (represented by a triangle), it’s evident that it cannot accurately capture the extremes of events like this one, where there are often multiple stations within a single cell. But that's not the end of the story. The model's “nominal” resolution—its grid spacing—is not a true reflection of the scale of phenomena it can resolve. To better understand this, think of pixels in an image: accurately detecting and representing the shape of an object requires a sufficient number of pixels. Similarly, for weather models, the “effective” resolution is typically 4 to 10 times coarser than the nominal resolution. In the plot, this limitation is illustrated by the larger triangles. As this event demonstrates, such highly localized phenomena are challenging to capture, even with relatively high resolutions. Accurately forecasting events like this requires increasingly high-resolution models with sub-kilometer grid spacing. However, even with such advancements, we remain constrained by uncertainties in the initial conditions, which propagate exponentially over time.

Good morning, Meteorologists and Atmospheric Scientists around the globe! Today, let's discuss why regional numerical weather models, such as the Weather Research and Forecasting (#WRF) model developed by NSF NCAR - The National Center for Atmospheric Research, are incredibly valuable to meteorologists worldwide. While global weather models like NOAA: National Oceanic & Atmospheric Administration's Global Forecast System (#GFS) and the European Centre for Medium-Range Weather Forecasts - ECMWF) model are widely utilized, they typically have grid spacing (resolution) of roughly 27km and 9km, respectively. Despite #ECMWF's finer resolution, it still falls within the convective "grey zone." This zone describes a modeling challenge where resolutions are too coarse to explicitly resolve convection yet too fine for convection to be adequately parameterized. Regional models like WRF address this limitation by taking the coarse resolution data from global models and downscaling it to finer grid spacing. This process significantly enhances forecast quality, providing more detailed and accurate representations of meteorological features. For instance, I recently conducted a WRF simulation over Côte d'Ivoire, driven by GFS data, using grids of 20km (to represent the approximate GFS native resolution) and 4km (convective-resolving scale). I've attached images highlighting the notable differences between these resolutions. First, consider the representation of topography. Due to its coarse grid spacing, the global model's resolution smooths out critical features such as river valleys, coastal inlets, and smaller #orographic details, potentially degrading forecast accuracy. In contrast, the high-resolution 4km WRF simulation clearly depicts these detailed terrain features. Second, let's examine precipitation forecasts. Both model resolutions can simulate precipitation totals effectively; however, the spatial distribution significantly differs. For example, the coarser 20km grid indicates an entire region near Dimbokro receiving uniform precipitation (e.g., 50-100mm). Meanwhile, the finer-scale 4km WRF model reveals a more nuanced and accurate distribution of rainfall across smaller areas, greatly improving the precision of forecasts. In summary, regional numerical models like WRF provide meteorologists with significantly enhanced spatial resolution, allowing for more detailed, accurate forecasts and better-informed weather predictions. These capabilities are essential for effective decision-making, particularly in areas sensitive to precise weather conditions. #Meteorology #WRF #WeatherModels #NumericalWeatherPrediction #AtmosphericScience #Forecasting #GFS #ECMWF #ConvectiveGreyZone #WeatherResearch

Attached screenshots show BBC Weather accurately predicting a brief shower around 12:30 PM on 10/07/2025, while MSN, AccuWeather, and WeatherWalay failed to indicate any rainfall at that time. Despite PMD forecasting rainfall over Gilgit-Baltistan from July 6–10, most days remained dry due to a dominant subtropical ridge suppressing vertical convection until its gradual breakdown around the 9th–10th. Most global weather apps (e.g., MSN, AccuWeather, WeatherWalay) rely on coarse-resolution models (GFS/ICON) that underperform in complex terrains like Gilgit. These models often miss short-lived, terrain-driven convection under weak monsoonal spillovers. In contrast, BBC Weather, powered by UK Met Office's higher-resolution Unified Model (UM), accurately captured localized rainfall potential. The gap between PMD’s broader window and actual rain timing reflects natural variability in synoptic-scale moisture surges and orographic triggers. Such micro-events highlight the need for high-res modeling and localized nowcasting in mountainous zones.

We're building a power system entirely dependent on the weather, and we're betting its stability on our ability to predict it. The transition to 100% renewable energy doesn't just change how we generate power. It changes what we need from meteorology. Wind and solar output are not dispatchable, they follow the atmosphere, not the market. This means grid operators, energy traders, and system planners now make critical decisions based on weather forecasts. And not just any forecast, they need predictions that are accurate across the right spatial scales and the right time horizons simultaneously. This is the domain of energy meteorology. And at its core is a forecasting chain that spans from global atmospheric models down to the turbine level, each scale serving a different decision, each with its own physics, resolution, and uncertainty. So how does a forecast actually get made? It starts with observation. Every hour, a global network of weather stations, radiosondes, ocean buoys, commercial aircraft, and satellites feeds roughly 10 million data points into numerical models. This isn't just data collection, it requires a process called data assimilation (4D-Var), which synchronizes real-world observations with the model's internal state across both space and time. Without it, the forecast drifts from reality before it even begins. This feeds the Global Weather Prediction (GWP) model, the backbone of the entire chain. It resolves large-scale atmospheric dynamics: cyclones, anticyclones, storm systems. Coarse resolution, high computational cost, but capable of forecasting up to 16 days ahead. Essential for long-term energy planning and capacity scheduling. But a global model can't see a valley, a coastal cliff, or a mountain pass. This is where mesoscale modelling takes over, zooming into regional domains, capturing local flow features like gap flows and orographic effects that directly shape wind resource variability at specific sites. And when you need to make decisions in the next few hours? That's the Rapid Refresh Model, high resolution, limited area, running every single hour. The operational heartbeat of intraday energy markets. Each scale feeds the next. Each serves a different decision horizon. But even the Rapid Refresh Model stops short of the turbine. At the microscale, we're no longer solving for weather, we're solving for energy. This is where atmospheric flow interacts with terrain roughness, forest canopies, building clusters, and ultimately, the rotor itself. The questions change: How does turbulence intensity vary with height? How does the boundary layer stratification shift between day and night? How does one turbine's wake reshape the wind seen by the next? This is the domain of tools like CFD, LES, and site-specific wind flow models, where the physics gets expensive and the stakes get specific. #WeatherForecasting #NumericalWeatherPrediction #EnergyMeteorology #AtmosphericScience