This paper explains how, but it's a tough slog. What follows is my summary.
The utility companies already solve a similar problem. Electricity is not easily stored, and in modern grids it is generally not stored at all. So, when you flick on a light switch, the immediate effect is that the power dissipated by all the lights in your neighborhood drops a little to compensate. Within seconds, some power turbine perhaps hundreds of miles away must twist a little harder on its generator shaft to get everyone's line voltage back up, and the extra thermal or hydroelectric power fed into the turbine to get this extra twist will be just about what your light bulb burns, plus the inefficiencies of getting the power from the turbine input to your bulb.
Turbines can only throttle up to 100% of their rated capacity, and they get inefficient when they throttle down too far, so utilities will shut down or spin up units to make larger changes. Changes like these take a long time, so utilities predict what the expected load at any given time will be, hours or days in advance, and schedule units to be on or off line to match the predicted load. Utilities keep some fraction of their turbines at partial output so that they can immediately crank up to match unexpected increases in the load.
The biggest increase in the load that they plan for is usually an unexpected dropout of one of the generators. If a 1 gigawatt generator suddenly goes off line, the grid controllers might respond by taking four other generators from 700 to 950 MW output. It would be impossible for this to happen instantly, but luckily, when most generators go offline, they do so gradually, and if they coast down over the course of seconds, other generators can crank up to match. If a circuit breaker pops or a line parts, or something else happens very quickly, then there is usually a temporary brownout as the line voltage drops down to the point where the loads match the generation. The backup generators usually ramp up within seconds, and many devices (like your computer or TV) can ride through a partial loss of power for a second or so.
So, the bottom line is that utilities predict the change in demand on their generators, and there is some variation from their prediction, and being ready for this variation costs money because some turbines (the spinning reserve) must be run at partial throttle which is less efficient than flat-out.
Just as an aside: consider how valuable it would be for the utility company to be able to instantly shut down your air conditioner for just a few minutes. This ability would act as part of their spinning reserve. During the summer, air conditioners are a substantial fraction of the total power burned. I'm pretty sure that for most of the U.S., the ability to shut down even a fraction of the air conditioners for 10 minutes would cover the entire spinning reserve requirement. That could save a lot of money, and PG&E (my local utility in California) is experimenting with just that through their Underfrequency Relay Option on the Base Interruptible Program. Anyway, back to the summary.
Wind farms produce electricity whenever the wind blows. Wind speeds can be predicted, and there is always variation from the prediction. When a wind farm is connected to the grid, the total variation in load on the load-following turbines is larger than without the wind farm, and so more turbines must be run in load-following mode, and these incur a cost associated with wind power that is real but not easy to predict before the wind farm is built.
For instance, part of Denmark's grid is connected to Norway's grid. Norway gets most of its power from hydroelectric plants. Hydroelectric plants are very good at load following and so they are usually the first choice of plant to handle variation from plan. Because Norway has lots of hydroelectric plants, and because it has high-throughput connections to Denmark, Denmark can hook up fairly high powered wind farms to its grid and incur relatively low costs for standby power.
Now that utilities are connecting large amounts of wind power to their grids, they are getting more precise numbers on the costs of doing so.
- Wind works well where you have year-round high winds near hydroelectric dams.
- The short-term variation from wind farms is usually quite small, since turbines are small (a few megawatts) and don't all shut off at the same time.
- Big storms give the worst case variation, since when a wind turbine goes too fast it feathers its blades and shuts down, going from full output to nothing, often in synchrony with other wind turbines around it.
- Wind farms spread over large geographic areas have less total variation (the wind doesn't die everywhere at the same time). Ideally the spinning reserve would thus scale up slower than the total windpower connected, making marginal wind power less expensive. Unfortunately Denmark is too small to see this effect, and it would require very high throughput long distance power distribution.
Wind turbines appear to work economically (when the utilities are prodded by a production tax credit, which I support). As more turbines are installed, the best windy areas are used up and the least expensive spinning reserve is committed. On the other hand, wind turbine costs might be coming down some day (maybe -- materials costs are going up), and the need for spinning reserve is decreasing. It's a fairly tense balance.
Personally, I'm happy to see more wind turbines getting installed, since it's domestically produced, mostly-carbon-free, low marginal cost power, and let's have more of that. There is going to be a limit to how much wind power can be installed, but we're nowhere near that limit yet.
At the same time, it's sobering to consider that the United States once built almost 100 nuclear plants in about two decades, bringing new power online at the average rate of 5 gigawatts a year, at a time when our economy was one-third to one-half the size it is now, in real terms. At it's peak, we were building much faster than that. This economic explosion was driven by the business fundamentals as much as it was by overexcited businessmen jumping on the latest bandwagon. And the fundamentals were and are that if we decide on a reactor design (like Palo Verde), we can build and operate them cheaper than coal plants.
To match this performance, and get to 20% of the U.S. grid in 20 years, the wind industry would need to install about 300 gigawatts of nameplate capacity. That would require getting to a peak of 30 gigawatts a year, up from 4 gigawatts in 2007, which is 11 years of sustained 20% growth. The fundamentals for wind are not as good as they were for nuclear in the 1960s. It won't happen without a major breakthrough.